Paper deep dive

MM-FinEval: A Multi-Task Multimodal Benchmark for Real-World Financial Forecasting

Dong Shu, Yanguang Liu, Huopu Zhang, Saisai Hu, Haiyan Zhao, Hekun Huang, Mengnan Du

Year: 2026Venue: arXiv preprintArea: cs.LGType: PreprintEmbeddings: 95

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%

Last extracted: 10/1/2026, 8:31:27 AM

Summary

The paper introduces MM-FinEval, a novel multi-task multimodal benchmark for real-world financial forecasting using earnings conference calls. The dataset contains 2,045 S&P 500 conference calls from 2019-2022, integrating three modalities: text transcripts, presentation slides, and audio recordings. It includes 12 financial task labels covering cumulative abnormal returns, stock volatility, and corporate bond returns. The authors evaluate 19 baseline models across Image-Text, Audio-Text, and Any-to-Any configurations, finding that small Any-to-Any models leveraging all three modalities outperform larger proprietary models restricted to fewer modalities, demonstrating the value of non-redundant multimodal information.

Entities (15)

MM-FinEval · datasetbenchmark · 99%Audio Recording · modality · 95%Presentation Slides · modality · 95%S&P 500 · indexentity · 95%Text Transcript · modality · 95%Any-to-Any Models · model-category · 90%Corporate Bond Return · financial-metric · 90%Cumulative Abnormal Return · financial-metric · 90%Directional Accuracy · metric · 90%Mean Absolute Error · metric · 90%Multimodal LLMs · model-category · 90%Stock Volatility · financial-metric · 90%Gemma 4 E2B · model · 85%GPT-5.4 · model · 85%Qwen 3.5 · model · 85%

Relation Signals (10)

MM-FinEval → containsmodality → Presentation Slides

confidence 95% · the corresponding presentation slides used during the call

MM-FinEval → containsmodality → Audio Recording

confidence 95% · and the entire audio recording.

MM-FinEval → containsmodality → Text Transcript

confidence 95% · Each input contains three modalities: a word-to-word text transcript of the earning call

Gemma 4 E2B → belongstocategory → Any-to-Any Models

confidence 90% · This includes Gemma 4 E2B and Gemma 4 E4B

MM-FinEval → evaluates → Multimodal LLMs

confidence 90% · MM-FinEval, a novel benchmark designed to evaluate multimodal LLMs across multiple financial tasks.

MM-FinEval → haslabel → Corporate Bond Return

confidence 90% · MM-FinEval is annotated with 12 labels... Corporate Bond Return

MM-FinEval → haslabel → Stock Volatility

confidence 90% · MM-FinEval is annotated with 12 labels... Volatility

MM-FinEval → haslabel → Cumulative Abnormal Return

confidence 90% · MM-FinEval is annotated with 12 labels... Cumulative Abnormal Return

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Financial forecasting from earnings conference calls requires models to reason over complex corporate disclosures, market expectations, and subtle communication signals. However, existing financial benchmarks are often limited to unimodal inputs or single-task settings, making it difficult to evaluate whether multimodal large language models (LLMs) can support real-world financial analysis. In this paper, we introduce MM-FinEval, a novel benchmark designed to evaluate multimodal LLMs across multiple financial tasks. MM-FinEval spans a diverse timeline from 2019 to 2022. The entire proposed dataset contains 2,045 S\&P 500 conference earning calls as inputs and 12 financial task labels as outputs. Each input contains three modalities: a word-to-word text transcript of the earning call, the corresponding presentation slides used during the call, and the entire audio recording. To establish a rigorous evaluation framework, we analyze 19 baseline models across three distinct model categories: Image-Text, Audio-Text, and Any-to-Any configurations. We observe that small-size Any-to-Any models processing all three modalities achieve strong performance, even when compared against larger proprietary models restricted to two-modality inputs. This indicates that our tri-modal dataset design introduces useful, non-redundant information. These results validate that text, audio, and visual data serve as important, complementary signals that mimic the decision-making process of expert human analysts.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

92,191 characters extracted from source content.

Expand or collapse full text

M-FinEval: A Multi-Task Multimodal Benchmark for Real-World Financial Forecasting Dong Shu 1 , Yanguang Liu 2 , Huopu Zhang 3 , Saisai Hu 4 , Haiyan Zhao 2 , Hekun Huang 2 , Mengnan Du 5 1 Northwestern University 2 New Jersey Institute of Technology 3 Georgia Institute of Technology 4 Pace University 5 The Chinese University of Hong Kong, Shenzhen Abstract Financial forecasting from earnings conference calls requires models to reason over complex corporate disclosures, market expectations, and subtle communication signals. However, exist- ing financial benchmarks are often limited to unimodal inputs or single-task settings, mak- ing it difficult to evaluate whether multimodal large language models (LLMs) can support real- world financial analysis. In this paper, we in- troduce M-FinEval, a novel benchmark de- signed to evaluate multimodal LLMs across multiple financial tasks. M-FinEval spans a diverse timeline from 2019 to 2022. The entire proposed dataset contains 2,045 S&P 500 conference earning calls as inputs and 12 financial task labels as outputs. Each input contains three modalities: a word-to-word text transcript of the earning call, the corresponding presentation slides used during the call, and the entire audio recording. To establish a rig- orous evaluation framework, we analyze 19 baseline models across three distinct model cat- egories: Image-Text, Audio-Text, and Any-to- Any configurations. We observe that small-size Any-to-Any models processing all three modal- ities achieve strong performance, even when compared against larger proprietary models re- stricted to two-modality inputs. This indicates that our tri-modal dataset design introduces useful, non-redundant information. These re- sults validate that text, audio, and visual data serve as important, complementary signals that mimic the decision-making process of expert human analysts. The M-FinEval dataset is available athttps://github.com/Tizzzzy/ M-FinEval. 1 Introduction Earnings conference calls are important events for public companies, serving as a primary channel for management to communicate financial perfor- mance, strategic outlooks, and operational updates to the market. The quality of a company’s earnings conference call is directly linked to the next-day stock market performance. In practice, financial analysts rely heavily on these events, extracting signals from company presentations and Q&A ses- sions to analyze corporate health and forecast fu- ture performance. Concurrently, there is a rapidly growing trend of leveraging LLMs to analyze fi- nancial markets and automate data processing (Wu et al., 2023; Li et al., 2023; Huang et al., 2023; Yang et al., 2023). If LLMs can be trained to suc- cessfully capture the dense, multi-layered informa- tion presented during these calls, financial institu- tions could benefit from an automated and scalable AI finance analysis pipeline that has historically required substantial human analysis. Despite the rapid advancement of LLMs, exist- ing financial models remain constrained by their reliance on uni-modal architectures and their ten- dency to focus on narrow, specific financial tasks (He and Zhang, 2025; Zhu et al., 2025; Liang and Carrasco Kind, 2025; Konstantinidis et al., 2024; Liu et al., 2025b). A major reason for this limita- tion is that the majority of current financial datasets are heavily skewed toward a single modality, such as pure text or numerical time-series data (Chen et al., 2021; Lai et al., 2025; Shah et al., 2023; Dong et al., 2024). Furthermore, even when recent datasets incorporate multimodal elements beyond text, they are typically restricted to a specific task (Shu et al., 2025a; Li et al., 2020; Luo et al., 2025; Shu et al., 2025b). This combination of unimodal and task-specific training data severely restricts financial LLMs to learn from additional contex- tual information, ultimately limiting their broader applicability in complex, real-world financial envi- ronments. To bridge this gap, we propose M-FinEval, a large-scale, multi-modal, and multi-label financial benchmark dataset designed specifically for earn- ings call analysis. Spanning a period from 2019 to 2022, the dataset encompasses a total of 2045 data 1 arXiv:2609.38523v1 [cs.LG] 29 Sep 2026 All Publicly Traded Firms in US Firm Collec t ion AAPL MSFT AMZN NVDA GOOGL KO WMT ... Audio Recordings PPT Slides Web Scrape Text Transcript Cumulative Abnormal Return End-of-Month Corporate Bond Return Last 5 Days of Month Corporate Bond Return Last Trading Day of Month Corporate Bond Return 3-Day Future Volatility 3-Day Pre-event Volatility 7-Day Future Volatility 7-Day Pre-event Volatility 15-Day Future Volatility 15-Day Pre-event Volatility 30-Day Future Volatility 30-Day Pre-event Volatility M M -FinEval L abels Oper ator : Good day exeryone ... Executives: This is a historic ... Analysts: I want to ask you ... Figure 1: M-FinEval Construction Pipeline. The dataset construction follows a three-stage process: assembling a foundational collection of publicly traded US firms, systematically web-scraping synchronized tri-modal data (text transcripts, audio recordings, and slides), and annotating the multimodal inputs with 12 financial task labels. points from S&P 500 companies, with 450 from 2019, 273 from 2020, 499 from 2021, and 823 from 2022. Each data point integrates three modal- ities: word-to-word text transcripts of the earnings call, the corresponding raw earnings call audio, and the accompanying presentation slide decks. Fur- thermore, M-FinEval is annotated with 12 labels, providing a benchmark to evaluate the performance of multimodal financial models. Through an ex- tensive experiment with 19 state-of-the-art models, we found that: •Current LLMs exhibit a severe performance gap between market return and volatility fore- casting. •Integrating all three modalities provides performance advantage, proving that non- redundant visual and audio information are essential for replicating expert financial rea- soning. • Model performance on volatility tasks aligns with the financial theory of information half- life. As the forecasting horizon expands from 3 to 30 days, the earnings signal is rapidly diluted by unpredictable market factors. 2 Related Work 2.1 Earnings Conference Call Datasets Earnings conference calls are important events for public companies because the quality of the infor- mation disclosed during these presentations and Q&A sessions is directly linked to next-day stock market performance. Recently, researchers have increasingly deployed machine learning models to systematically interpret the information in these calls (Sawhney et al., 2020b,a; Mathur et al., 2022; Sawhney et al., 2020c), aiming to uncover quantita- tive investment signals and making money. This ef- fort to uncover quantitative investment signals has motivated the community to publish specialized datasets focusing on earnings conference calls. For example, ECTSum (Mukherjee et al., 2022) pro- vides benchmarks for bullet-point summarization of highly verbose transcripts. FinNLI (Magomere et al., 2025) is developed to assess logical deduc- tion. SubjECTive-QA (Pardawala et al., 2024) mea- sures the subjectivity of Q&A responses across conference earning calls. Recent large text datasets like STRUX (Lu et al., 2024) and DEC (Yu et al., 2025) created for LLM-driven investment decision- making. However, these existing datasets remain con- strained by their reliance on unimodal architectures, do not include the acoustic and visual signals in- herent in the actual conference call. EC (Qin and Yang, 2019) which was the first dataset to combine textual transcripts with raw audio recordings to pre- dict financial risk. Building upon this, MAEC (Li et al., 2020) scaled the available conference call data and optimized to predict stock volatility. Dat- sets such as Earnings-21 (Del Rio et al., 2021) and ConEC (Huang et al., 2024) focus heavily on the acoustic challenges of the domain, serving primar- ily as benchmarks for automatic speech recognition rather than broader financial reasoning. FinCall- Surprise (Shu et al., 2025a), which provides con- ference call text transcripts, audio, and slides, is optimized exclusively for predicting earnings sur- prises. Despite the success of these existing multi- modal conference call datasets, they are limited to a specific financial task. This severely restricts the ability of financial LLMs to learn from addi- tional contextual information, directly motivating 2 the need for multi-label multi-modal benchmarks. 3 M-FinEval Construction The construction of M-FinEval dataset follows a three-stage pipeline designed to collect, synchro- nize, and annotate data from multiple sources. As illustrated in Figure 1, the first stage involves gath- ering a set of liquid public firms in the United States. To ensure data quality, we restrict the sam- ple to companies with a market capitalization ex- ceeding $1 billion and an average daily trading volume above $50 million. This initial screening process yielded a pool of over 4,000 companies. In the second stage, we acquire data across the three modalities for each firm’s earnings confer- ence calls. We systematically web-scrape these contents from official company websites. The col- lected text transcripts are highly structured, featur- ing clear speaker identifiers (operators, executives, and analysts) alongside their corresponding state- ments. Because most publicly traded firms host one earnings call per fiscal quarter, we ensure temporal consistency by manually aligning all data sources according to their specific reporting dates. Fur- thermore, we cross-reference the conference call titles to verify that the transcript, audio recording, and presentation slides all originate from the exact same event, ensuring synchronization across our multimodal dataset. 3.1 Multi-Label Preliminaries M-FinEval benchmark includes 12 labels de- signed to predict a company’s cumulative abnormal return, stock volatility, and corporate bond returns based on its quarterly earnings call. The complete labels are shown on the right side of Figure 1. Cumulative Abnormal Return.Cumulative ab- normal return (CAR) measures the market-adjusted price reaction around an earnings call event (Hirsh- leifer et al., 2009; Duan et al., 2018). For each firm iand event datet, we first compute the abnormal return by subtracting a benchmark market return from the firm’s realized return: AR i,t = R i,t − R m,t ,(1) whereR i,t denotes the stock return of firmion day t, and R m,t denotes the corresponding market return. We then aggregate abnormal returns over an event window [t 1 ,t 2 ]: CAR i,[t 1 ,t 2 ] = t 2 X t=t 1 AR i,t .(2) CAR reflects whether the information revealed in the earnings call leads to a positive or negative market reassessment of the firm. We include CAR- based labels over multiple post-call windows to evaluate whether models can infer short-term and medium-term directional market reactions from multimodal earnings-call information. Volatility.Volatility measures stock-price fluctu- ations over a finite period and is commonly used as a measure of risk (Kogan et al., 2009). In our study, the period is defined relative to the earnings conference call using pre-event and post-event win- dows of different lengths. Following Kogan et al. (Kogan et al., 2009), volatility from dayt− τto day t is defined as v [t−τ,t] = r P τ i=0 (r t−i − ̄r) 2 τ ,(3) wherer t is the return on dayt, ̄ris the mean return over the period of dayt − τto dayt. r t = P t /P t−1 − 1is the return price, withP t denot- ing the closing price on dayt. While CAR captures the direction of the market response, volatility cap- tures stock-price instability regardless of whether returns are positive or negative. The labelsfut_3d– fut_30dandpast_3d–past_30dmeasure post- event and pre-event volatility, respectively, over windows of 3, 7, 15, and 30 calendar days. Corporate Bond Return. Corporate bond re- turns capture the monthly performance of a firm’s outstanding bonds and provide a debt-market mea- sure of investors’ responses to information dis- closed during earnings calls. The bond return data are observed at the monthly level and are matched to conference-call observations using the firm’s stock ticker, recorded asCOMPANY_SYMBOL, and the bond trading month, recorded inDATE. Specifically, monthly returns are computed as the percentage change in the bond’s dirty price (clean price plus accrued interest), adjusted for coupon payments received during the holding period, following the standard methodology in the corporate bond lit- erature (Bessembinder et al., 2008; Jostova et al., 2013). To account for differences in bond trading frequency and liquidity, we consider three comple- mentary return measures that differ only in their 3 Transcript (words)Slide (pages)Audio (sec) YearMiniMaxAvg.MiniMaxAvg.MiniMaxAvg.Total 20192477154218837.90115625.971096.078870.973642.66450 20202533218949389.90117027.891404.067585.313859.25273 20212784153438998.85111327.471197.9511938.893720.29499 2022 2986177178842.89225230.73234.7231393.573770.45823 Table 1: Data statistics of the M-FinEval. The dataset spans four years (2019–2022), with each conference call containing three synchronized modalities: text transcripts, presentation slides, and audio recordings. We report the minimum, maximum, and average values for each modality. Transcript length is measured in words, slides in pages, and audio in seconds. The last column shows how much data we have for each year. definition of the month-end price. RET_EOM is calculated using the last observed transaction price during the calendar month, regardless of the trading date. RET_LDM uses the transaction price on the last trading day of the month and is missing if the bond does not trade on that day. RET_L5M uses the last transaction price observed within the final five trading days of the month, providing a balance between month-end pricing and sample coverage. 4 Data Analysis Table 1 presents a detailed statistical analysis of our dataset, summarizing its scale and multimodal structure. M-FinEval has a total of 2045 data spanning from 2019 to 2022. On average, an in- dividual data features a text transcript of approxi- mately 9,000 words, accompanied by a 27-page presentation slide deck and an audio recording spanning roughly 3,700 seconds. As illustrated in Figure 2a, the content of these calls is con- sistently structured around three primary speaker roles: Operators, Executives, and Analysts. Op- erators typically manage the flow of the event by opening and closing the call. Executives, such as the CEO and CFO, deliver the company’s financial results and strategic outlook. Finally, Analysts rep- resenting investment firms and financial institutions ask questions to obtain more information into the company’s operational and financial performance. More data analysis including our label statistic is in Appendix A. 5 Benchmark Settings 5.1 Baseline LLMs To establish a comprehensive performance bench- mark for M-FinEval, we selected 19 state-of-the- art open- and closed-source models. These models were categorized into three architectural groups based on their input capabilities. The first group comprises Large Vision- Language Models (LVLMs), which process both image and text inputs. This group includes the Qwen 3.5 series (spanning 2B to 35B parameters) (Qwen Team, 2026a), the Qwen 3.6 series (27B and 35B) (Qwen Team, 2026b), Gemma 3n (Team et al., 2024), and the Mistral 3 series (3B, 8B, and 14B) (Jiang et al., 2023). It also includes proprietary models: GPT-5.4, GPT-5.5, Claude 4.6 Sonnet, and Claude 4.8 Opus. The second group consists of Text-Audio Language Models, which process both audio and text inputs, represented in our study by Voxtral 24B (Liu et al., 2025a) and Moss 8B (Team, 2026b). Finally, the third group features Any-to-Any Models, which are natively equipped to process all modalities, text, audio, and images, simultaneously. This includes Gemma 4 E2B and Gemma 4 E4B (Team, 2026a). 5.2 Benchmark Input Design To systematically evaluate the performance of the baseline models, we designed a customized input and prompting strategy tailored to each modality. The full set of prompts used in the experiments is provided in Appendix B. Image-Text Modality:Because the selected baseline models represent the current state-of-the- art, they all support extended context windows ex- ceeding 128K tokens. Given that the longest text transcript in our dataset is 21,894 words, we input the raw transcripts directly without applying any preprocessing steps, such as textual summariza- tion. However, the visual data required structural adaptation. The original presentation slides are collected in PDF format. Since most models do not natively support PDF input, we converted each 4 (a) Transc ript Example(b) Labels St at ist ic L abels:mean std min max car: 0.003443 0.054573 -0.261699 0.312476 ret_eom: 0.001929 0.024981 -0.208935 0.154950 ret_l5m: 0.001877 0.025233 -0.208935 0.132135 ret_ldm: 0.001620 0.025573 -0.207922 0.121361 fut_3d: 0.017588 0.014283 0.000439 0.152479 fut_7d: 0.019605 0.011929 0.002890 0.114500 fut_15d: 0.019874 0.011657 0.004166 0.143844 fut_30d: 0.020287 0.011600 0.004408 0.161032 past_3d: 0.020148 0.015334 0.000288 0.133771 past_7d: 0.020292 0.011701 0.002662 0.100579 past_15d: 0.019352 0.009554 0.004090 0.074227 past_30d:0.019842 0.010549 0.004775 0.092727 I nput: Operator: Good day, everyone. Welcome to today's conference and webcast to discuss .... Executives: Thank you, John. Good morning, everyone. This is a historic day for both .... Analysts: Giovanni, I want to ask you where the cell therapy platform fits in your .... Operator: We will now go live to the presentation. Please stand by... Executives: I'm going to take you through the earnings presentation, which is available... Analysts: Just a question on loan growth, 2% to 3% on an annual average basis... ... Figure 2: Transcript Structure and Label Statistics. page into a separate image. To accommodate the image-count limits inherent to many of these mod- els, we implemented a random sampling strategy. For each presentation, we randomly selected five images from the slide deck, explicitly excluding the first and last pages, which typically contain non-substantive content such as title cards or con- cluding remarks. In cases where a presentation contained five or fewer substantive slides, all avail- able slides were used. The final input sequence was structured as: “Images Instruction Text Tran- script Answer Format”. Although this specific benchmark experiment restricts the visual input to five sampled images, the complete PDF files are included in the public dataset release. Audio-Text Modality:We structured the input sequence as: “Audio Instruction Text Tran- script Answer Format”. Image-Audio-Text Modality:To ensure a fair comparison across all baseline models, we main- tained the exact input constraints established in the previous setups. We used the same randomly se- lected slides as the Image-Text setup. The final input sequence for this comprehensive modality was structured as: “Images Audio Instruc- tion Text Transcript Answer Format”. 5.3 Benchmark Metrics Because all of our labels are related to stock re- turns or volatility, the ground truth data consists of continuous decimal values. Therefore, we frame the prediction as a regression task. In this context, Mean Absolute Error (MAE) serves as the default metric for evaluating absolute model performance on our benchmark. For a dataset ofnsamples, MAE is defined as: MAE = 1 n n X i=1 |y i − ˆy i |,(4) wherey i represents the ground-truth numerical label andˆy i denotes the model’s prediction. How- ever, given the inherent complexities of financial forecasting, relying on a single evaluation metric is insufficient to capture a model’s true reasoning ca- pabilities. Therefore, we also employ the Pearson Correlation Coefficient (r), formulated as: r = P n i=1 (y i − ̄y)(ˆy i − ̄ ˆy) q P n i=1 (y i − ̄y) 2 P n i=1 (ˆy i − ̄ ˆy) 2 .(5) This metric evaluates the model’s capacity to correctly identify the relative magnitude of market reactions across a broad cross-section of compa- nies. Finally, we report Directional Accuracy (DA). In financial markets, accurately forecasting the di- rection of a movement is a important signal for investment strategies. DA measures the proportion of predictions that correctly anticipate the mathe- matical sign of the actual outcome, expressed as: DA = 1 n n X i=1 1(sgn(y i ) = sgn(ˆy i )),(6) where1is the indicator function andsgnextracts the positive or negative sign of the value. 5 LabelMetric Image-TextAudio-TextAny-to-Any Qw3.5-2BQw3.5-4BQw3.5-9BQw3.5-27BQw3.5-35BQw3.6-27BQw3.6-35BGm3nMis3-3BMis3-8BMis3-14BGPT-5.4GPT-5.5Cld-4.6Cld-4.8Vox-24BMossGm4-E2BGm4-E4B car MAE↓0.05110.04980.04990.05340.04700.05050.04780.07400.05990.05400.05460.04930.05260.04930.04580.04290.10790.04180.0437 DA↑21.76%41.61%48.07%50.56%53.11%51.64%51.10%55.88%46.11%50.66%51.98%51.83%53.45%53.69%41.09%20.79%37.60%55.17%53.20% ret_eom MAE↓0.03410.01980.02100.03020.01990.02280.01960.03460.01840.01840.01840.01820.01440.01790.01770.02243.42770.02970.0248 DA↑37.85%25.28%51.00%56.23%55.55%56.77%57.11%56.65%53.55%56.82%56.33%59.49%71.46%71.92%72.41%21.71%37.16%60.59%60.10% ret_l5m MAE↓0.03160.01960.02070.02210.01920.02010.01920.03350.01930.02460.01990.01820.01710.01910.01920.02100.07480.02290.0216 DA↑26.50%26.45%52.96%56.67%55.35%56.09%54.03%56.16%5.77%51.25%54.72%59.22%66.01%67.00% 67.49%23.47%40.78%59.11%61.08% ret_ldm MAE↓0.03180.02060.02060.02150.01880.01950.01870.03530.01980.02250.01970.01830.01800.01880.01940.02310.13690.02950.0234 DA↑25.23%34.23%51.10%55.21%54.38%53.74%56.63%55.17%6.45%53.69%54.87%59.22%64.46%67.98%66.50%27.09%39.76%62.56%62.07% fut_3d MAE↓0.02300.02150.02620.02530.01690.02140.01540.03050.02610.03190.03080.02520.02210.01420.01390.019221.71550.01150.0089 DA↑47.04%95.11%96.09%96.58%99.41%96.33%99.32%97.54%71.44%97.75%97.26%99.76% 100.00%100.00%100.00%35.60%53.64%98.03%99.51% fut_7d MAE↓0.02350.01900.02360.02270.01570.01910.01560.02980.02670.07490.03180.02540.02420.01500.01060.01960.66680.01020.0087 DA↑46.11%95.89%95.84%95.84%99.22%97.11%98.68%100.00%66.21%96.43%97.36%100.00%100.00%99.01%100.00%32.08%53.50%98.52%99.51% fut_15d MAE↓0.02380.01980.02410.02180.01630.02050.02540.02830.02420.12920.03440.02650.02670.0157 0.00970.01940.66330.00970.0085 DA↑46.50%95.89%95.94%98.48%99.27%98.34%99.22%98.03%51.64%94.67%97.26%100.00%100.00%97.54%100.00%24.01%53.30%98.52%99.34% fut_30d MAE↓0.02460.02150.02550.0227 0.01660.05740.14950.03310.02570.20640.05730.02870.02500.05860.01910.01980.13380.01060.0086 DA↑54.67%95.50%96.72%97.90%98.78%98.09%98.58%98.52%58.68%94.96%97.21%100.00%100.00%98.03%100.00%31.39%52.81%98.03%99.19% past_3d MAE↓0.02060.02130.03310.02290.01770.01800.02190.03970.03930.10950.03370.01500.01310.01180.01230.02290.18230.01720.0152 DA↑4.01%81.08%95.26%99.17%95.45%98.88%84.16%93.60%53.64%95.31%99.32%100.00%100.00%100.00%100.00%47.33%56.87%94.58%55.17% past_7d MAE↓0.02050.02060.03470.02140.01530.01380.02880.04270.05090.24980.04140.01420.01660.01430.01140.02260.71530.02070.0140 DA↑3.42%81.12%95.06%99.66%98.04%99.32%89.19%96.06%48.95%95.35%99.27%100.00%100.00%100.00%100.00%46.70%56.48%91.63%57.14% past_15d MAE↓0.02090.02060.03440.02140.01500.01570.06680.04820.04840.27860.05460.01380.01770.03400.01180.02160.75780.04030.0125 DA↑8.85%81.17%95.06%99.66%99.12%99.46%89.63%93.43%32.57%95.84%99.36%100.00%100.00%100.00%100.00%47.48%57.21%92.61%63.05% past_30d MAE↓0.02280.02120.03740.03560.01860.11990.21200.06050.05640.28720.17530.02450.02430.10680.06020.02411.12580.07420.0120 DA↑15.99%78.78%94.72%99.61%99.46%99.32%94.72%95.07%35.94%95.75%98.24%100.00%100.00%100.00%100.00%50.71%56.48%91.63%69.46% Table 2: Overall Performance of 19 Baseline Models on M-FinEval. The evaluation reports averaged Mean Absolute Error (MAE) and Directional Accuracy (DA) across 12 financial tasks spanning the 2019–2022 period. For each row, the highest score for each metric is highlighted in a distinct color corresponding to the model type. 5.4 Implementation Details All experiments were conducted on two NVIDIA A100 GPUs, each with 80GB of memory. For all baseline models, we maintained their official repos- itory configurations and loaded them with bfloat16 precision to optimize computational efficiency. Be- yond this precision setting, no other modifications were made to the models’ default parameters, en- suring a fair and reproducible comparison across all baseline evaluations. 6 Benchmark Results 6.1 The Gap Between Returns and Volatility As shown in Table 2, an initial analysis of the benchmark results reveals a large gap in the ability of LLMs to predict market returns versus market volatility. Across the majority of baseline models, Directional Accuracy (DA) for return-based tasks (car, ret_eom, ret_l5m, ret_ldm) is relatively poor, frequently ranging between 40% and 60%. Only few models achieve above 70% (e.g., Claude-4.8 on ret_eom). Because DA evaluates only the binary sign of the prediction, a score hovering near 50% indicates that these models only marginally outper- form random guessing. In contrast, the majority of models achieve a DA over 90%, and occasionally reach 100%, when predicting volatility. Exceptions are rare, with only a few small-size models, such as Qwen3.5-2B, falling below 50%. However, as illustrated by the label statistics in Figure 2b, it is common knowledge in the financial domain that volatility cannot be negative. Therefore, despite achieving high DA scores on volatility-based tasks, small-size LLMs still fail to preserve a basic un- derstanding of fundamental constraints in financial domains. This performance gap is further illuminated by evaluating the Mean Absolute Error (MAE). For instance, high-performing models (e.g., Claude, Gemma-4) achieve MAEs between 0.041 and 0.045 on the Cumulative Abnormal Return (car) task. As shown in Figure 2b, return-based tasks are inher- ently noisy: their ground truth means hover near zero while exhibiting high variance. Consequently, this MAE is large compared to car ground truth mean of 0.003443 and representing approximately 77% of the dataset’s standard deviation. In contrast, volatility tasks demonstrate a tighter error distri- bution. On tasks like fut_15d, top models achieve MAEs as low as 0.0085 to 0.0097, which represent a much smaller fraction of the ground truth mean (0.019874) and standard deviation (0.011657). We present a detailed breakdown of each model’s an- nual MAE performance in Table 3, capturing the granular trends over each consecutive year from 2019-2022. 6.2 Inconsistent Scaling Laws and the Proprietary Advantage Within specific model architectures and tasks, our benchmark in Table 2 reveals a clear scaling law where increased parameter count directly corre- lates with enhanced financial reasoning capabili- 6 Qw3.5-2B Qw3.5-4B Qw3.5-9B Qw3.5-27B Qw3.5-35B Qw3.6-27B Qw3.6-35B Gm3n Mis3-3B Mis3-8B Mis3-14B GPT-5.4 GPT-5.5 Cld-4.6 Cld-4.8 Vox-24B Moss Gm4-E2B Gm4-E4B car ret_eom ret_l5mret_ldm fut_3dfut_7d fut_15d fut_30dpast_3dpast_7d past_15dpast_30d 0.0 -0.1 0.1 0.2 0.3 0.4 0.5 Figure 3: Overall Pearson Correlation Coefficients of Baseline Models. The heatmap illustrates the models’ capacity to recognize linear trends in financial data. ties. This trend is particularly pronounced within the Qwen 3.5 and 3.6 model family. As the pa- rameter size scales from 2 billion to 35 billion, we observe a consistent, upward trajectory in overall Directional Accuracy across both return-based and volatility tasks. However, this scaling law breaks down on the MAE metric. Rather than exhibiting a consistent decrease as parameter counts grow, the models’ MAE changes unpredictably across tasks. For ex- ample, within the Qwen 3.5 family on the Cumula- tive Abnormal Return task, the 27B model yields an MAE of 0.0534, performing worse than its much smaller 4B (0.0498) and 9B (0.0499) counter- parts. Similar inconsistencies appear across other model families. For instance, the Mistral-3B model achieves a lower MAE (0.0261) on the fut_3d task than both the Mistral-8B (0.0319) and Mistral-14B (0.0308). These findings suggest that while larger models have a better ability to predict the market movement direction, they do not have the knowl- edge to accurately estimate the magnitude of those movements. Beyond model family scaling, the benchmark results highlight a performance gap between open- source and proprietary models. Within the Image- Text modality, the GPT and Claude model families consistently outperform their open-source coun- terparts. Leading proprietary models, such as GPT-5.5 and Claude-4.8, achieve the highest tier of Directional Accuracy across the majority of tasks. Specifically, they reach DA scores of ap- proximately 72% on ret_eom, and 67% on both ret_l5m and ret_ldm. These results outperform the near-random guessing baseline (approximately 50%) in return based tasks. Furthermore, their 100% DA on volatility tasks suggests that these advanced proprietary models have a basic under- standing of the non-negativity constraints inherent to financial volatility. 6.3 The Advantage of Any-to-Any Multimodal Integration As shown in Table 2, our benchmark results show the advantage of integrating all three data modali- ties: text, audio, and visual. The Any-to-Any mod- els evaluated in this study, specifically Gemma 4 E2B and Gemma 4 E4B, perform well on both return-based and volatility tasks. Despite their small size architectures (Gemma 4 E2B is a 5B parameter model utilizing only 2B active param- eters, while Gemma 4 E4B is an 8B model with 4B active parameters) both achieve a Directional Accuracy exceeding 60% across ret_eom, ret_l5m, 7 0.6 0.4 0.2 0.0 -0.2 -0.4 -0.6 2019202020212022201920202021202220192020202120222019202020212022 fut_15dfut_30dfut_3dfut_7d past_3dpast_7dpast_15dpast_30d Qw3.5-2B Qw3.5-4B Qw3.5-9B Qw3.5-27B Qw3.5-35B Qw3.6-27B Qw3.6-35B Gm3n Mis3-3B Mis3-8B Mis3-14B GPT-5.4 GPT-5.5 Cld-4.6 Cld-4.8 Vox-24B Moss Gm4-E2B Gm4-E4B Qw3.5-2B Qw3.5-4B Qw3.5-9B Qw3.5-27B Qw3.5-35B Qw3.6-27B Qw3.6-35B Gm3n Mis3-3B Mis3-8B Mis3-14B GPT-5.4 GPT-5.5 Cld-4.6 Cld-4.8 Vox-24B Moss Gm4-E2B Gm4-E4B Figure 4: Annual Pearson Correlation Trends (2019–2022) on Volatility Tasks. The heatmap illustrates how models’ ability to capture linear relationships shifts over time. and ret_ldm. This performance is better than much larger Image-Text modality baselines, such as the 35B parameter Qwen 3.6 model, and even propri- etary GPT 5.4 model. Furthermore, on the Cumu- lative Abnormal Return task, both Gemma 4 E2B (0.0418) and Gemma 4 E4B (0.0437) achieve lower MAEs than the highly capable Claude-4.8 model (0.0458). This unexpected performance strongly validates that textual, visual, and audio data streams provide distinct, complementary financial signals rather than redundant information. In real-world finan- cial markets, institutional analysts do not evaluate corporate disclosures solely based on the confer- ence call transcript. They synthesize what man- agement is presenting via visual slides, the literal lexical content of the transcript, and, critically, how that content is delivered, extracting subtle acoustic cues from vocal tone, pacing, and inflection. To accurately predict market reactions, LLMs must develop a similarly holistic comprehension. The superior performance of these Any-to-Any mod- els shows that equipping LLMs with this complete spectrum of human communication is essential for replicating expert-level financial reasoning. 6.4 Conservative Versus Aggressive Prediction in Audio-Text Models Another interesting observation in Table 2 is that, in Audio-Text columns, Voxtral consistently achieves a superior MAE across the entire benchmark, whereas Moss performs better on DA. A closer examination reveals that Voxtral exhibits highly conservative prediction behaviors, frequently gen- erating its output close to zero. By minimizing the absolute magnitude of its predictions, Voxtral successfully avoids absolute error penalties, result- ing in a relatively lower MAE. However, this zero- anchored conservatism prevents the model from capturing meaningful market momentum, leading to exceptionally poor DA scores. Conversely, Moss adopts a highly aggressive predictive strategy. It frequently generates big numerical prediction, and is more likely to capture the correct directional sign. 6.5 Pearson Correlation In addition to MAE and DA, we evaluated the Pear- son correlation coefficient to assess the models’ ca- pacity to recognize linear trends and relative shifts in the financial data. As illustrated in Figure 3, open-source models generally exhibit low Pear- son correlations across all tasks, with the highest scores peaking near 0.20 on the vol_fut_15d task (achieved by Qwen3.5 27B and Gemma 3n). Inter- estingly, proprietary models such as GPT-5.4, GPT- 5.5, and the Claude 4 series achieve surprisingly moderate-to-strong positive Pearson correlations on volatility tasks. For instance, GPT-5.4 reaches a correlation of 0.44 on vol_fut_7d, and Claude 4.8 reaches 0.42. This results reveals that proprietary architectures can successfully detect relative shifts 8 car ret_eom ret_l5m past_3dpast_7dpast_15dpast_30d fut_3d fut_7d 2019202020212022 0 40 80 0 40 80 0 40 80 ret_ldm fut_15dfut_30d 201920202021202220192020202120222019202020212022 Figure 5: Annual Directional Accuracy (DA) Trends Across M-FinEval (2019–2022). Each line represents an individual model’s performance over the four-year period, illustrating how predictive accuracy shifts across different financial tasks. in market risk. We also observe a universal failure across all architectures to predict Cumulative Abnormal Re- turn (car). While proprietary models demonstrate strong positive correlations on discrete return met- rics, with GPT-5.5 achieving an impressive 0.60 on ret_eom (End of Month Return). Both open-source and closed-source models exhibit near-zero or even negative correlations on the CAR task. This dis- crepancy likely stems from the structural complex- ity of Cumulative Abnormal Return, which isolates the firm-specific stock reaction by mathematically regressing against broader market benchmarks over a specific event window. Finally, the correlation data exposes a clear tem- poral degradation in model performance across volatility prediction horizons. For the proprietary models, predictive capability is highest at the short- term 3-day and 7-day intervals, but gradually de- cays through the 15-day mark, ultimately col- lapsing at the 30-day horizon for both post-event and pre-event volatility. This progressive decay strongly aligns with the financial theory of infor- mation half-life: the qualitative and quantitative signals extracted directly from an earnings call are most potent in the immediate short term (Hafez et al., 2021; Molinaro, 2026). As the prediction window expands to a full month, the initial catalyst of the earnings signal is rapidly diluted and over- whelmed by unpredictable market factors. This confirms that forecasting long-term variance is in- herently unfeasible for current LLM architectures. To further validate the information half-life the- ory, we map the Pearson correlation coefficients for the volatility tasks across the 2019–2022 time- line in Figure 4. This visualization reinforces the theory by revealing a clear temporal degradation as the prediction horizon expands. For instance, when observing the proprietary models across post- event volatility windows, GPT and Claude achieve strong correlations for the 3-day and 7-day intervals (fut_3d and fut_7d); however, this predictive capa- bility steadily deteriorates at the 15-day and 30-day marks (fut_15d and fut_30d). An isolated excep- tion occurs with Claude 4.8, which exhibits unusu- ally high performance across all post-event volatil- ity tasks specifically in the year 2021. Despite this outlier, the overarching trend of progressive decay strongly aligns with our findings in Figure 3, reaffirming that the predictive signal contained within a corporate disclosure is rapidly diluted by unpredictable market factors over extended tempo- ral windows. More results analysis is in Appendix C. 9 6.6 Temporal Degradation on Direction Accuracy To evaluate how model performance changes over time, we analyze the annual Directional Accuracy trends from 2019 to 2022. As illustrated in Figure 5, there is a consistent decline in Directional Ac- curacy from 2019 to 2022 across the return-based tasks (ret_eom, ret_l5m, and ret_ldm). In 2019, several top-performing models tightly cluster near peak accuracies between 70% and 90%. However, this performance decreases and drops to a DA range of 30% to 40% in 2022. This highlights a lack of temporal generalization in current LLMs. It sug- gests that post-pandemic market dynamics intro- duced increased market complexity that reduced the models’ predictive capabilities. 7 Conclusion In this work, we introduced M-FinEval, a multi- modal benchmark for evaluating LLMs across 12 financial tasks using 2,045 S&P 500 earnings con- ference calls spanning a 2019–2022 period. The three modalities inputs in M-FinEval reflect the analysis process of expert human financial analysts, who must synthesize not only what executives dis- close but how they present it to execute complex financial reasoning. Our extensive baseline evalu- ation across Image-Text, Audio-Text, and Any-to- Any configurations exposes the boundary where pure language modeling must transition into multi- sensory, calibrated quantitative reasoning. We hope this dataset serves as a benchmark for the AI and finance communities, supporting future work to- ward multimodal financial analysis for economic forecasting. 10 References Hendrik Bessembinder, Kathleen M Kahle, William F Maxwell, and Danielle Xu. 2008. Measuring ab- normal bond performance. The Review of Financial Studies, 22(10):4219–4258. Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan R Routledge, and 1 others. 2021. Finqa: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697–3711. Miguel Del Rio, Natalie Delworth, Ryan Wester- man, Michelle Huang, Nishchal Bhandari, Joseph Palakapilly, Quinten McNamara, Joshua Dong, Pi- otr Zelasko, and Miguel Jetté. 2021. Earnings-21: A practical benchmark for asr in the wild. arXiv preprint arXiv:2104.11348. Zihan Dong, Xinyu Fan, and Zhiyuan Peng. 2024. Fn- spid: A comprehensive financial news dataset in time series. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Min- ing, pages 4918–4927. Junwen Duan, Yue Zhang, Xiao Ding, Ching Yun Chang, and Ting Liu. 2018. Learning target-specific representations of financial news documents for cu- mulative abnormal return prediction. In Proceed- ings of the 27th international conference on compu- tational linguistics, pages 2823–2833. Peter Hafez, Ricard Matas Navarro, Francisco Gomez, and Marko Kangrga. 2021. Three trading signals generated systematically from earnings conference calls. Available at SSRN 4323558. Zhenzhi He and Cong Zhang. 2025. Cross-sectional spillovers of earnings surprises and asset price anoma- lies. Available at SSRN 5415255. David Hirshleifer, Sonya Seongyeon Lim, and Siew Hong Teoh. 2009. Driven to distraction: Ex- traneous events and underreaction to earnings news. The Journal of Finance, 64(5):2289–2325. Allen H Huang, Hui Wang, and Yi Yang. 2023. Finbert: A large language model for extracting information from financial text. Contemporary Accounting Re- search, 40(2):806–841. Ruizhe Huang, Mahsa Yarmohammadi, Jan Trmal, Jing Liu, Desh Raj, Leibny Paola Garcia, Alexei Ivanov, Patrick Ehlen, Mingzhi Yu, Ariya Rastrow, Dan Povey, and Sanjeev Khudanpur. 2024. ConEC: Earnings call dataset with real-world contexts for benchmarking contextual speech recognition. In Pro- ceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3700– 3706, Torino, Italia. ELRA and ICCL. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. Preprint, arXiv:2310.06825. Gergana Jostova, Stanislava Nikolova, Alexander Philipov, and Christof W Stahel. 2013. Momentum in corporate bond returns. The Review of Financial Studies, 26(7):1649–1693. Shimon Kogan, Dimitry Levin, Bryan R Routledge, Ja- cob S Sagi, and Noah A Smith. 2009. Predicting risk from financial reports with regression. In Pro- ceedings of human language technologies: the 2009 annual conference of the North American Chapter of the Association for Computational Linguistics, pages 272–280. Thanos Konstantinidis, Giorgos Iacovides, Mingxue Xu, Tony G Constantinides, and Danilo Mandic. 2024. Finllama: Financial sentiment classification for algorithmic trading applications. arXiv preprint arXiv:2403.12285. Viet Lai, Michael Krumdick, Charles Lovering, Varshini Reddy, Craig Schmidt, and Chris Tanner. 2025. Sec- qa: A systematic evaluation corpus for financial qa. In Proceedings of The 10th Workshop on Financial Technology and Natural Language Processing, pages 221–236. Jiazheng Li, Linyi Yang, Barry Smyth, and Ruihai Dong. 2020. Maec: A multimodal aligned earnings confer- ence call dataset for financial risk prediction. In Pro- ceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 3063–3070. Yinheng Li, Shaofei Wang, Han Ding, and Hang Chen. 2023. Large language models in finance: A survey. In Proceedings of the fourth ACM international con- ference on AI in finance, pages 374–382. Qingwen Liang and Matias Carrasco Kind. 2025. How does managers’ willingness to disclose affect an- alysts’ earning forecasts-a measurement by llms. Available at SSRN 5199752. Alexander H Liu, Andy Ehrenberg, Andy Lo, Clé- ment Denoix, Corentin Barreau, Guillaume Lam- ple, Jean-Malo Delignon, Khyathi Raghavi Chandu, Patrick von Platen, Pavankumar Reddy Muddireddy, and 1 others. 2025a.Voxtral.arXiv preprint arXiv:2507.13264. Zhaowei Liu, Xin Guo, Zhi Yang, Fangqi Lou, Lingfeng Zeng, Jinyi Niu, Mengping Li, Qi Qi, Zhiqiang Liu, Yiyang Han, and 1 others. 2025b.Fin-r1: A large language model for financial reasoning through reinforcement learning.arXiv preprint arXiv:2503.16252. 11 Yiming Lu, Yebowen Hu, Hassan Foroosh, Wei Jin, and Fei Liu. 2024. Strux: An llm for decision- making with structured explanations. arXiv preprint arXiv:2410.12583. Junyu Luo, Zhizhuo Kou, Liming Yang, Xiao Luo, Jinsheng Huang, Zhiping Xiao, Jingshu Peng, Chengzhong Liu, Jiaming Ji, Xuanzhe Liu, and 1 oth- ers. 2025. Finmme: Benchmark dataset for financial multi-modal reasoning evaluation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29465–29489. Jabez Magomere, Elena Kochkina, Samuel Mensah, Simerjot Kaur, and Charese Smiley. 2025. FinNLI: Novel dataset for multi-genre financial natural lan- guage inference benchmarking. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 4545–4568, Albuquerque, New Mexico. Association for Computational Linguistics. Puneet Mathur, Mihir Goyal, Ramit Sawhney, Ritik Mathur, Jochen L Leidner, Franck Dernoncourt, and Dinesh Manocha. 2022. Docfin: Multimodal fi- nancial prediction and bias mitigation using semi- structured documents. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 1933–1940. Matéo Molinaro. 2026. Do earnings call transcripts predict post-announcement returns? Available at SSRN 6695758. Rajdeep Mukherjee, Abhinav Bohra, Akash Banerjee, Soumya Sharma, Manjunath Hegde, Afreen Shaikh, Shivani Shrivastava, Koustuv Dasgupta, Niloy Gan- guly, Saptarshi Ghosh, and Pawan Goyal. 2022. ECT- Sum: A new benchmark dataset for bullet point sum- marization of long earnings call transcripts. In Pro- ceedings of the 2022 Conference on Empirical Meth- ods in Natural Language Processing, pages 10893– 10906, Abu Dhabi, United Arab Emirates. Associa- tion for Computational Linguistics. Huzaifa Pardawala, Siddhant Sukhani, Agam Shah, Veer Kejriwal, Abhishek Pillai, Rohan Bhasin, Andrew DiBiasio, Tarun Mandapati, Dhruv Adha, and Sud- heer Chava. 2024. Subjective-qa: Measuring sub- jectivity in earnings call transcripts’ qa through six- dimensional feature analysis. Advances in Neural Information Processing Systems, 37:59342–59372. Yu Qin and Yi Yang. 2019. What you say and how you say it matters: Predicting stock volatility using verbal and vocal cues. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 390–401. Qwen Team. 2026a. Qwen3.5: Towards native multi- modal agents. Qwen Team. 2026b. Qwen3.6-35B-A3B: Agentic cod- ing power, now open to all. Ramit Sawhney, Arshiya Aggarwal, Piyush Khanna, Puneet Mathur, Taru Jain, and Rajiv Ratn Shah. 2020a. Risk forecasting from earnings calls acoustics and network correlations. In INTERSPEECH, pages 2307–2311. Ramit Sawhney, Piyush Khanna, Arshiya Aggarwal, Taru Jain, Puneet Mathur, and Rajiv Shah. 2020b. Voltage: Volatility forecasting via text audio fusion with graph convolution networks for earnings calls. In Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 8001–8013. Ramit Sawhney, Puneet Mathur, Ayush Mangal, Piyush Khanna, Rajiv Ratn Shah, and Roger Zimmermann. 2020c. Multimodal multi-task financial risk forecast- ing. In Proceedings of the 28th ACM international conference on multimedia, pages 456–465. Agam Shah, Suvan Paturi, and Sudheer Chava. 2023. Trillion dollar words: A new financial dataset, task & market analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 6664–6679. Dong Shu, Yanguang Liu, Huopu Zhang, and Mengnan Du. 2025a. Fincall-surprise: A large scale multi- modal benchmark for earning surprise prediction. arXiv preprint arXiv:2510.03965. Dong Shu, Haoyang Yuan, Yuchen Wang, Yanguang Liu, Huopu Zhang, Haiyan Zhao, and Mengnan Du. 2025b. Finchart-bench: Benchmarking finan- cial chart comprehension in vision-language models. arXiv preprint arXiv:2507.14823. Gemma Team. 2026a.Gemma 4 technical report. Preprint, arXiv:2607.02770. Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, and 1 others. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295. OpenMOSS Team. 2026b.Moss-audio techni- cal report.https://github.com/OpenMOSS/ MOSS-Audio. GitHub repository. Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kam- badur, David Rosenberg, and Gideon Mann. 2023. Bloomberggpt: A large language model for finance. arXiv preprint arXiv:2303.17564. Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. 2023. Fingpt: Open-source financial large language models. arXiv preprint arXiv:2306.06031. Ding Yu, Zhuo Liu, and Hangfeng He. 2025. Same com- pany, same signal: The role of identity in earnings call transcripts. In Findings of the Association for Computational Linguistics: ACL 2025, pages 18403– 18422. 12 Yu Zhu, Xiao Liu, and Olivia R Liu Sheng. 2025. Post- earnings-announcement drift prediction: Leveraging postevent investor responses with multitask learning. Information Systems Research, 36(4):2191–2212. 13 A Data Analysis In Figure 2, we detail the label statistics, including the mean, standard deviation, minimum, and max- imum for all 12 financial task labels. An analysis of these distributions explicitly confirms the struc- tural differences between discrete return tasks and continuous volatility tasks. The return-based tasks (car, ret_eom, ret_l5m, and ret_ldm) exhibit a classic zero-centered finan- cial return distribution. The means for these labels hover barely above zero, ranging from 0.001620 to 0.003443. In contrast, their standard deviations are substantially larger, ranging from 0.024981 to 0.054573. This high variance relative to the mean confirms a notoriously low signal-to-noise ratio in- herent to predicting market direction. Furthermore, the minimum and maximum values demonstrate that returns fluctuate wildly in both directions. For instance, the Cumulative Abnormal Return (car) spans from a minimum of -0.261699 to a maximum of 0.312476, illustrating the extreme, bidirectional stochastic shocks that models must account for. Conversely, the volatility-based tasks display a fundamentally different statistical profile. Most no- tably, the minimum values across all eight volatility labels are strictly positive (e.g., the lowest absolute minimum is 0.000288 for past_3d). This directly re- flects the hard non-negativity constraint of financial volatility. The means are significantly higher and tightly clustered between 0.017588 and 0.020292. Additionally, the standard deviations (0.009554 to 0.015334) are much smaller relative to their re- spective means when compared to the return labels. This points to a tighter, more structured distribution characterized by volatility clustering, making it a more autoregressive and stable phenomenon. These distinct distributional characteristics math- ematically explain the performance divergence ob- served in our benchmark results. Return tasks are heavily obscured by symmetric, zero-centered noise, making them highly resistant to accurate quantitative regression. Volatility tasks, however, present a structured, bounded, and clustered target that is quantitatively easier for models to map, per- fectly mirroring established financial econometrics. B Prompt Used in Our Experiments B.1 CAR Task Image-Text Models You are an expert quantitative equity analyst. Your task is to predict the stock’s Cumulative Abnormal Return (CAR) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the Cumulative Abnormal Return (CAR) asso- ciated with this earnings event. CAR measures the sum of the differences between the stock’s actual return and its expected return over the event window. 2. Provide 1-2 sentences explaining the key driver. 3. Your CAR prediction must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ CAR decimal ]] Audio-Text Models You are an expert quantitative equity analyst. Your task is to predict the stock’s Cumulative Abnormal Return (CAR) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the Cumulative Abnormal Return (CAR) asso- ciated with this earnings event. CAR measures the sum of the differences between the stock’s actual return and its expected return over the event window. 2. Provide 1-2 sentences explaining the key driver. 3. Your CAR prediction must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ CAR decimal ]] Any-to-Any Models You are an expert quantitative equity analyst. Your task is to predict the stock’s Cumulative Abnormal Return (CAR) for a company based on its earnings call transcript, earn- ings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the Cumulative Abnormal Return (CAR) asso- ciated with this earnings event. CAR measures the sum of the differences between the stock’s actual return and its expected return over the event window. 2. Provide 1-2 sentences explaining the key driver. 3. Your CAR prediction must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ CAR decimal ]] 14 B.2 ret_eom Image-Text Models You are an expert quantitative fixed-income analyst. Your task is to predict the End-of-Month corporate bond re- turn (ret_eom) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the ret_eom (End-of-Month corporate bond return). It represents the percentage gain or loss of an investment or business value over the course of one month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_eom prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_eom decimal ]] Audio-Text Models You are an expert quantitative fixed-income analyst. Your task is to predict the End-of-Month corporate bond re- turn (ret_eom) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the ret_eom (End-of-Month corporate bond return). It represents the percentage gain or loss of an investment or business value over the course of one month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_eom prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_eom decimal ]] Any-to-Any Models You are an expert quantitative fixed-income analyst. Your task is to predict the End-of-Month corporate bond re- turn (ret_eom) for a company based on its earnings call transcript, earnings call audio, and the accompanying pre- sentation slides. Transcript: text_transcript Instructions: 1. Predict the ret_eom (End-of-Month corporate bond return). It represents the percentage gain or loss of an investment or business value over the course of one month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_eom prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_eom decimal ]] B.3 ret_l5m Image-Text Models You are an expert quantitative fixed-income analyst. Your task is to predict the corporate bond return for the last 5 days of the month (ret_l5m) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the ret_l5m (Return Last 5 Days of Month). It represents the percentage gain or loss of the corporate bond during the final 5 trading days of the month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_l5m prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_l5m decimal ]] Audio-Text Models You are an expert quantitative fixed-income analyst. Your task is to predict the corporate bond return for the last 5 days of the month (ret_l5m) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the ret_l5m (Return Last 5 Days of Month). It represents the percentage gain or loss of the corporate bond during the final 5 trading days of the month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_l5m prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_l5m decimal ]] Any-to-Any Models You are an expert quantitative fixed-income analyst. Your task is to predict the corporate bond return for the last 5 days of the month (ret_l5m) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the ret_l5m (Return Last 5 Days of Month). It represents the percentage gain or loss of the corporate bond during the final 5 trading days of the month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_l5m prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_l5m decimal ]] 15 B.4 ret_ldm Image-Text Models You are an expert quantitative fixed-income analyst. Your task is to predict the corporate bond return specifically on the last trading day of the month (ret_ldm) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the ret_ldm (Return Last Day of Month). It represents the percentage gain or loss of the corporate bond exactly on the final trading day of the month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_ldm prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_ldm decimal ]] Audio-Text Models You are an expert quantitative fixed-income analyst. Your task is to predict the corporate bond return specifically on the last trading day of the month (ret_ldm) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the ret_ldm (Return Last Day of Month). It represents the percentage gain or loss of the corporate bond exactly on the final trading day of the month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_ldm prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_ldm decimal ]] Any-to-Any Models You are an expert quantitative fixed-income analyst. Your task is to predict the corporate bond return specifically on the last trading day of the month (ret_ldm) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the ret_ldm (Return Last Day of Month). It represents the percentage gain or loss of the corporate bond exactly on the final trading day of the month. 2. Provide 1-2 sentences explaining the key driver. 3. Your ret_ldm prediction must be represented as a deci- mal (e.g., 0.052 or -0.014). **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ ret_ldm decimal ]] B.5 vol_fut_3d Image-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 3-day volatility (vol_fut_3d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_3d (Volatility Future 3 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 3 trading days immediately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_3d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_3d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 3-day volatility (vol_fut_3d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the vol_fut_3d (Volatility Future 3 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 3 trading days immediately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_3d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_3d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to predict the post-event 3-day volatility (vol_fut_3d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_3d (Volatility Future 3 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 3 trading days immediately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_3d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_3d decimal ]] 16 B.6 vol_fut_7d Image-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 7-day volatility (vol_fut_7d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_7d (Volatility Future 7 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 7 trading days immediately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_7d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_7d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 7-day volatility (vol_fut_7d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the vol_fut_7d (Volatility Future 7 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 7 trading days immediately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_7d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_7d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to predict the post-event 7-day volatility (vol_fut_7d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_7d (Volatility Future 7 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 7 trading days immediately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_7d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_7d decimal ]] B.7 vol_fut_15d Image-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 15-day volatility (vol_fut_15d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_15d (Volatility Future 15 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 15 trading days imme- diately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_15d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_15d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 15-day volatility (vol_fut_15d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the vol_fut_15d (Volatility Future 15 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 15 trading days imme- diately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_15d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_15d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to predict the post-event 15-day volatility (vol_fut_15d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_15d (Volatility Future 15 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 15 trading days imme- diately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_15d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_15d decimal ]] 17 B.8 vol_fut_30d Image-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 30-day volatility (vol_fut_30d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_30d (Volatility Future 30 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 30 trading days imme- diately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_30d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_30d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to predict the post-event 30-day volatility (vol_fut_30d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Predict the vol_fut_30d (Volatility Future 30 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 30 trading days imme- diately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_30d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_30d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to predict the post-event 30-day volatility (vol_fut_30d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Predict the vol_fut_30d (Volatility Future 30 Days). This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 30 trading days imme- diately following this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_fut_30d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_fut_30d decimal ]] B.9 vol_past_3d Image-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 3-day volatility (vol_past_3d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_3d (Volatility Past 3 Days). This represents the realized volatility (price turbulence and un- certainty) of the asset over the 3 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_3d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_3d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 3-day volatility (vol_past_3d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Estimate the vol_past_3d (Volatility Past 3 Days). This represents the realized volatility (price turbulence and un- certainty) of the asset over the 3 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_3d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_3d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to estimate the pre-event 3-day volatility (vol_past_3d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_3d (Volatility Past 3 Days). This represents the realized volatility (price turbulence and un- certainty) of the asset over the 3 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_3d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_3d decimal ]] 18 B.10 vol_past_7d Image-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 7-day volatility (vol_past_7d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_7d (Volatility Past 7 Days). This represents the realized volatility (price turbulence and un- certainty) of the asset over the 7 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_7d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_7d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 7-day volatility (vol_past_7d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Estimate the vol_past_7d (Volatility Past 7 Days). This represents the realized volatility (price turbulence and un- certainty) of the asset over the 7 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_7d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_7d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to estimate the pre-event 7-day volatility (vol_past_7d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_7d (Volatility Past 7 Days). This represents the realized volatility (price turbulence and un- certainty) of the asset over the 7 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_7d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_7d decimal ]] B.11 vol_past_15d Image-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 15-day volatility (vol_past_15d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_15d (Volatility Past 15 Days). This represents the realized volatility (price turbulence and uncertainty) of the asset over the 15 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_15d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_15d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 15-day volatility (vol_past_15d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Estimate the vol_past_15d (Volatility Past 15 Days). This represents the realized volatility (price turbulence and uncertainty) of the asset over the 15 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_15d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_15d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to estimate the pre-event 15-day volatility (vol_past_15d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_15d (Volatility Past 15 Days). This represents the realized volatility (price turbulence and uncertainty) of the asset over the 15 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_15d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_15d decimal ]] 19 B.12 vol_past_30d Image-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 30-day volatility (vol_past_30d) for a company based on its earnings call transcript and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_30d (Volatility Past 30 Days). This represents the realized volatility (price turbulence and uncertainty) of the asset over the 30 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_30d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_30d decimal ]] Audio-Text Models You are an expert quantitative analyst. Your task is to estimate the pre-event 30-day volatility (vol_past_30d) for a company based on its earnings call transcript and earnings call mp3 audio. Transcript: text_transcript Instructions: 1. Estimate the vol_past_30d (Volatility Past 30 Days). This represents the realized volatility (price turbulence and uncertainty) of the asset over the 30 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_30d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_30d decimal ]] Any-to-Any Models You are an expert quantitative analyst. Your task is to estimate the pre-event 30-day volatility (vol_past_30d) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides. Transcript: text_transcript Instructions: 1. Estimate the vol_past_30d (Volatility Past 30 Days). This represents the realized volatility (price turbulence and uncertainty) of the asset over the 30 trading days immediately preceding this earnings call. 2. Provide 1-2 sentences explaining the key driver. 3. Your vol_past_30d estimate must be represented as a decimal. **Answer format** Rationale = [[ 1-2 sentences rationale ]] Prediction = [[ vol_past_30d decimal ]] C Pearson Correlation Temporal Observations We evaluated the Pearson correlation coefficient across the 2019–2022 timeline. Figure 6 indicates that proprietary models improve their ability to cap- ture linear relationships over time. For example, on the ret_eom task, GPT-5.5 exhibits a strict up- ward trajectory, scaling from a correlation of 0.309 in 2019 to 0.581 by 2022. Similarly, Claude 4.8 improves from 0.144 in 2019 to 0.437 in 2022 on the exact same task. This temporal strengthening suggests that massive, closed-source models are highly adept at mapping evolving macroeconomic environments to relative asset rankings. In contrast, open-source models demonstrate no such temporal adaptation. Models like Qwen 3.5 and Mistral 3 remain near-zero correlations on ret_eom across the four-year span. D Examples of Model Responses Qwen3.5 2B Rationale = The Cumulative Abnormal Return (CAR) is predicted to be positive, driven by the company’s achieve- ment of all-time record quarterly and full-year operating income and operating ratio, which significantly outper- formed analyst expectations and the broader market. Prediction = [[ 0.045 ]] Qwen3.5 4B Rationale = [[ The company reported a 22% decline in GAAP earnings per share due to significant one-time costs, including severance charges and litigation expenses, which negatively impacted the market’s sentiment despite under- lying non-GAAP growth. The management’s focus on cost-cutting and the uncertainty surrounding future market conditions likely led to a negative market reaction.]] Prediction = [[ -0.035 ]] Qwen3.5 9B Rationale = The company delivered solid full-year results with revenue growth, margin expansion, and strong free cash flow, but the stock price likely already priced in the positive momentum from the second half of the year. Ad- ditionally, the guidance for 2019 EPS of $13.90 was pre- sented as a ̈ floor ̈ rather than a target, and the company ex- plicitly noted that the Red Hat acquisition (a major value driver) is excluded from 2019 financial expectations due to timing uncertainty, which may limit the upside surprise in the immediate reaction. Prediction = [[ 0.012 ]] Qwen3.5 27B Rationale = [[ The company reported a 22% year-over- year decline in GAAP earnings per share to $0.84, missing expectations due to significant one-time charges for sever- 20 0.6 0.4 0.2 0.0 -0.2 -0.4 -0.6 2019202020212022201920202021202220192020202120222019202020212022 fut_15dfut_30dfut_3dfut_7d past_3dpast_7dpast_15dpast_30d carret_eom ret_l5mret_ldm Qw3.5-2B Qw3.5-4B Qw3.5-9B Qw3.5-27B Qw3.5-35B Qw3.6-27B Qw3.6-35B Gm3n Mis3-3B Mis3-8B Mis3-14B GPT-5.4 GPT-5.5 Cld-4.6 Cld-4.8 Vox-24B Moss Gm4-E2B Gm4-E4B Qw3.5-2B Qw3.5-4B Qw3.5-9B Qw3.5-27B Qw3.5-35B Qw3.6-27B Qw3.6-35B Gm3n Mis3-3B Mis3-8B Mis3-14B GPT-5.4 GPT-5.5 Cld-4.6 Cld-4.8 Vox-24B Moss Gm4-E2B Gm4-E4B Qw3.5-2B Qw3.5-4B Qw3.5-9B Qw3.5-27B Qw3.5-35B Qw3.6-27B Qw3.6-35B Gm3n Mis3-3B Mis3-8B Mis3-14B GPT-5.4 GPT-5.5 Cld-4.6 Cld-4.8 Vox-24B Moss Gm4-E2B Gm4-E4B Figure 6: Annual Pearson Correlation Trends (2019–2022) on the M-FinEval Benchmark. The heatmap illustrates how models’ ability to capture linear relationships shifts over time. ance, litigation, and headquarters relocation totaling $0.16 per share. Although non-GAAP earnings grew 9%, the market reaction is likely to be negative due to the headline GAAP miss and the explicit guidance that equity markets will not be a significant contributor to revenue growth in 2019. ]] Prediction = [[ -0.035 ]] Qwen3.5 35B Rationale = PNC delivered a strong 2020 with record revenue and positive operating leverage, but the market reaction is likely tempered by the guidance for a modest decline in Net Interest Income (NII) for 2021 and the significant drag on reserves ($3.2 billion provision) due to the pandemic, resulting in a neutral to slightly negative market response. Prediction = -0.012 Qwen3.6 27B Rationale = [[ The company significantly raised its full- year 2021 revenue guidance to a range of $5.5-$6.0 billion and projected a substantial expansion in operating margins to 10.4%-11.0%, driven by strong net order growth and a robust backlog. Additionally, the firm highlighted a strong balance sheet with a low debt-to-capital ratio and a clear path to double-digit community count growth, signaling strong future profitability. ]] Prediction = [[ 0.065 ]] Qwen3.6 35B Rationale = BlackRock reported record full-year revenue growth of 20% and operating income growth of 19%, sig- nificantly outperforming expectations with record net in- flows of $540 billion and a 10% increase in AUM to $10 trillion. The company also announced a strong capital return program, including an 18% dividend increase and $1.5 billion in share repurchases, signaling high confi- dence in future cash flows and growth. Prediction = [[ 0.035 ]] Gemma3n Rationale = [[ The earnings call presented a strong Q3 performance with solid growth across key products like Opdivo and Eliquis, and positive momentum from new launches like Zeposia and Abecma. The company reaf- firmed its financial guidance and expressed confidence in its ability to navigate the challenges of generic erosion while continuing to invest in innovation, which is likely to be viewed favorably by investors. ]] Prediction = [[ 0.06 ]] 21 LabelYear Image-TextAudio-TextAny-to-Any Qw3.5-2BQw3.5-4BQw3.5-9BQw3.5-27BQw3.5-35BQw3.6-27BQw3.6-35BGm3nMis3-3BMis3-8BMis3-14BGPT-5.4GPT-5.5Cld-4.6Cld-4.8Vox-24BMossGm4-E2BGm4-E4B car 20190.05310.04860.04950.05250.04700.04930.04590.06180.05880.05270.05210.04890.05220.04500.04750.04370.11310.04100.0418 20200.05210.05020.05170.05430.04850.05410.05140.05940.06000.05690.05650.05070.05220.04350.04950.04510.09510.03310.0344 20210.04710.04810.04690.05160.04410.04680.04450.10170.05900.05080.05230.04570.04950.06200.04480.04150.10300.05010.0536 20220.05200.05220.05130.05530.04850.05170.04930.07320.06170.05580.05730.05170.05490.04660.04140.04110.12050.04300.0449 ret_eom 20190.03600.02030.02110.03050.02020.02450.01950.03120.01860.01840.01850.01830.01420.01780.01770.02291.00740.02410.0240 20200.03920.01930.02100.03150.01940.02460.02140.04110.01780.01780.01780.01760.01350.01700.01730.02151.44560.02320.0225 20210.02980.02010.02100.03070.02040.02170.01910.03740.01870.01870.01870.01850.01500.01820.01780.02333.61730.02300.0226 20220.03140.01950.02110.02790.01950.02060.01850.02870.01830.01870.01870.01850.01480.01850.01800.02171.91410.04830.0302 ret_l5m 20190.03370.02010.02080.02220.01970.02040.01940.02950.01910.02420.02000.01840.01660.01920.01920.02140.08270.01780.0209 20200.03420.01910.02070.02240.01930.02060.01970.03980.01880.02390.02050.01770.01630.01810.01840.02000.08980.01710.0195 20210.02790.01990.02080.02220.01950.01990.01900.03600.01940.02530.01960.01840.01740.01940.01940.02150.06380.02170.0213 20220.03040.01940.02060.02140.01850.01930.01850.02870.01980.02520.01960.01850.01820.01950.01970.02110.06300.03490.0247 ret_ldm 20190.03420.02130.02080.02180.01920.02000.01900.03150.01980.02200.01970.01850.01760.01910.01950.02380.15850.02130.0225 20200.03450.02030.02080.02210.01880.02030.01950.04250.01970.02220.02050.01780.01710.01790.01860.02210.16010.02130.0212 20210.02800.02080.02060.02160.01890.01940.01870.03810.01960.02280.01940.01860.01820.01900.01950.02410.11180.03050.0235 20220.03040.02000.02040.02060.01810.01840.01770.02920.02010.02320.01920.01850.01920.01900.02000.02260.11710.04500.0264 fut_3d 20190.02140.02100.02500.02350.01520.01950.01380.02530.02580.03220.03040.02470.02050.01250.01230.01841.34120.01010.0075 20200.02590.02460.03150.02840.02010.02500.01980.04220.03050.03680.03350.02980.02610.01780.01630.022748.06400.01420.0107 20210.02260.02110.02480.02560.01620.02170.01410.03160.02410.02890.02980.02340.02180.01350.01410.01770.15830.01160.0084 20220.02210.01940.02340.02350.01600.01930.01400.02300.02410.02960.02930.02290.02010.01300.01280.017837.29850.01010.0090 fut_7d 20190.02150.01860.02250.02130.01430.01750.01410.02510.02640.07680.03140.02490.02270.01340.00920.01880.61250.00910.0076 20200.02660.02190.02830.02550.01810.02250.01930.04090.03050.08050.03450.02980.02860.01830.01240.02330.81260.01250.0104 20210.02330.01850.02220.02270.01490.01900.01410.03110.02460.06820.03060.02360.02330.01440.01070.01770.61520.01020.0083 20220.02270.01710.02120.02120.01560.01750.01500.02220.02540.07400.03070.02320.02220.01390.01010.01850.62710.00890.0084 fut_15d 20190.02180.01940.02300.02050.01500.01890.02600.02420.02420.12930.03400.02600.02520.01420.00850.01860.61050.00880.0076 20200.02680.02250.02890.02430.01870.02390.02700.03820.02710.13880.03710.03100.03120.01910.01130.02300.80370.01180.0099 20210.02350.01930.02280.02180.01550.02060.02260.02980.02230.12170.03330.02470.02590.01480.00970.01760.61110.00960.0082 20220.02310.01810.02170.02060.01590.01840.02590.02090.02300.12710.03320.02420.02460.01470.00930.01850.62790.00840.0083 fut_30d 20190.02250.02100.02420.02130.01530.05430.14930.02850.02580.20790.05710.02810.02380.05920.01750.01900.13110.00960.0076 20200.02780.02420.03050.02520.01900.05780.14850.04370.02840.21950.06070.03350.02870.05780.02180.02330.13960.01290.0101 20210.02440.02100.02400.02280.01590.05790.15040.03450.02390.19830.05600.02670.02460.05930.01880.01800.13100.01060.0084 20220.02370.01970.02310.02140.01630.05970.15010.02560.02450.20000.05530.02630.02290.05790.01810.01880.13370.00930.0083 past_3d 20190.01930.02100.03150.02240.01740.01790.02140.03220.03960.11180.03280.01480.01310.01180.01210.02230.19700.01630.0151 20200.02340.02420.04010.02700.02060.02100.02570.05430.04350.12070.03910.01760.01530.01390.01450.02700.22010.02050.0175 20210.01980.02040.03060.02160.01670.01700.02090.04160.03710.10300.03160.01400.01230.01090.01150.02130.15170.01660.0142 20220.01970.01950.03010.02050.01620.01610.01970.03080.03690.10260.03130.01360.01180.01040.01110.02110.16040.01540.0140 past_7d 20190.01920.02040.03300.02100.01510.01380.02840.03510.05140.25200.04060.01410.01660.01440.01130.02210.65590.01960.0139 20200.02310.02340.04160.02530.01750.01600.03260.05810.05580.27210.04680.01660.01910.01660.01330.02660.86470.02570.0161 20210.01990.01960.03230.02020.01450.01310.02750.04540.04870.23720.03920.01330.01570.01330.01070.02100.66220.01920.0131 20220.01990.01900.03180.01920.01410.01230.02660.03220.04770.23790.03880.01290.01520.01290.01030.02060.67840.01830.0127 past_15d 20190.01950.02040.03270.02110.01490.01570.06650.03960.04940.28040.05370.01370.01780.03400.01170.02110.69700.03810.0124 20200.02340.02340.04100.02550.01710.01820.07170.06450.05310.30130.06050.01610.02020.03740.01360.02530.90800.04910.0144 20210.02030.01970.03210.02010.01430.01500.06500.05040.04630.26620.05220.01290.01670.03280.01100.02010.70220.03780.0118 20220.02040.01890.03170.01890.01380.01410.06410.03810.04480.26670.05190.01250.01620.03200.01060.01990.72400.03600.0113 past_30d 20190.02140.02100.03540.03550.01850.12020.21180.04960.05730.28820.17410.02440.02450.10680.05940.02371.05440.07060.0119 20200.02570.02410.04410.04160.02130.12460.21730.07920.06170.30900.18340.02870.02780.11180.06730.02821.30910.08740.0140 20210.02210.02030.03500.03340.01770.11800.21000.06430.05430.27550.17180.02300.02300.10510.05770.02231.05600.07090.0113 20220.02210.01940.03500.03210.01700.11660.20880.04890.05250.27620.17180.02220.02210.10370.05630.02211.08380.06790.0109 Table 3: Annual MAE Trends (2019–2022) on the M-FinEval Benchmark. Mistral3 3B Rationale = [The strong execution in 2018, with signifi- cant improvements in ROTCE (14.1%), efficiency ratio (56.7%), and underlying EPS growth, alongside disci- plined balance sheet growth and robust credit quality, sug- gests strong investor confidence. The company’s strategic initiatives like TOP programs, balance sheet optimiza- tion, and digital transformation, combined with a pos- itive economic outlook and prudent risk management, should drive positive abnormal returns. The stock’s cur- rent valuation discount relative to peers, especially in a recession-sensitive sector, and the company’s ability to sustain growth through organic and strategic acquisitions, are likely to support a positive CAR.] Prediction = [0.065] Mistral3 8B Rationale =Capital One’s fourth quarter 2019 earnings call and presentation highlighted several positive yet nuanced financial developments. The company reported strong loan growth (7% year-over-year), solid revenue growth (7%), and a stable credit environment, though offset by increased provisions for credit losses (up 11%) and the transition to CECL, which added $2.9B to the allowance. The Walmart partnership launch, digital transformation progress, and moderate marketing expense growth suggest sustained growth momentum, but the CECL impact and higher provisioning could dampen near-term profitability expectations. The net charge-off rate remained stable at 1.93%, and the company’s disciplined capital management and strategic investments in technology and marketing indicate long-term value creation. Prediction = [[ 0.025 ]]", Mistral3 14B Rationale = [[BNY Mellon’s earnings call highlighted strong organic growth, innovation in digital assets and real-time payments, and a robust capital position, which should positively surprise investors. The guidance for double-digit EPS growth in 2022, driven by higher net interest revenue and fee revenue recovery, further supports a positive market reaction.]] Prediction = [[0.035]] 22 GPT 5.4 Rationale = [[ Kimberly-Clark delivered a clear 2020 beat with strong EPS, margin expansion, cash flow, and im- proved market share, but the market reaction is likely tempered by a cautious 2021 outlook featuring only 1-2% organic growth and significant commodity inflation that is expected to pressure margins. The net effect should be modestly positive because execution and restructuring/cost savings were strong, yet guidance implies a tougher earn- ings quality and margin backdrop ahead. ]] Prediction = [[ 0.012 ]] GPT 5.5 Rationale = [[ Despite record full-year results, the event was dominated by investor concern over a Q4 EPS miss driven by sharply higher compensation/non-comp ex- penses, wage inflation, and weaker sequential trading performance. Management’s commentary on continued investment spending, capital headwinds, and limited near- term buybacks likely reinforced negative revisions to for- ward profitability. ]] Prediction = [[ -0.065 ]] Claude 4.6 Rationale = [[ PMI delivered a strong beat with Q2 ad- justed diluted EPS of $1.57 (+17.8% organic), raised full- year EPS guidance to $5.97-$6.07 (implying +15-17% in dollar terms), confirmed a $5-7B share buyback program, and announced the imminent IQOS ILUMA Japan launch all meaningfully positive catalysts that likely exceeded consensus expectations and drove a positive abnormal re- turn. ]] Prediction = [[ 0.038 ]] Claude 4.8 Rationale = [[ WBA delivered a strong Q1 beat (adjusted EPS $1.68, up 53%, well ahead of expectations) driven by COVID vaccinations/testing and robust retail comps, plus raised full-year guidance, which should be received posi- tively despite headwinds from shrink, labor investments, and softer underlying scripts. ]] Prediction = [[ 0.045 ]] Voxtral 24B **Answer** Rationale = The company’s strong financial performance, with a 32% increase in comparable basis EPS excluding canopy equity earnings, and the announcement of a $2 billion share repurchase program, are the key drivers for the stock’s CAR. The company’s commitment to returning $5 billion in value to shareholders through fiscal 23 also supports a positive outlook. Prediction = 0.035 Moss Rationale = [[ This earnings call seems to be a ’surprising positive’ event which can lead to a strong positive CAR. It may provide a signal that the outlook for the company is improving, leading to an increase in the company’s stock price. ]] Prediction = 0.123 Gemma4 E2B Rationale = [[ The company reported strong retail sales growth in Q3, driven by improved mix and digital sales, and provided guidance indicating a strong Q3 and a cau- tious outlook for Q4. The financial results show a positive trend in retail sales and margin expansion, which are posi- tive indicators for the stock. ]] Prediction = [[ 0.025 ]] Gemma4 E4B Rationale = [[ The earnings call presents a positive out- look with strong underlying sales growth and a significant increase in adjusted segment EBIT, which should lead to a positive market reaction. The company also high- lights strong free cash flow and a positive ESG momentum, which further supports a bullish sentiment. ]] Prediction = [[ 0.035 ]] 23

Instant research discovery