Paper deep dive
FinVerse: Financial Time-Series Benchmark
Jaehoon Lee, Jun Seo, Seunghan Lee, Tae Yoon Lim, Dongwan Kang, Hwanil Choi, Minjae Kim, Sungdong Yoo, Junhyeok Kang, Sangjun Han, Soonyoung Lee, Wonbin Ahn
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As time-series foundation models have emerged, the need for benchmarks that can evaluate their forecasting ability in meaningful ways has become increasingly important. Existing time-series forecasting benchmarks provide useful standardized comparisons, but they often evaluate heterogeneous series with uniform error-based metrics. Strong performance under such metrics does not necessarily imply that a model's forecasts will support the best real-world decisions across domains. For example, in stock forecasting, correctly predicting whether a price will rise or fall can be more directly relevant to realized returns than minimizing point-wise forecast error alone. To this end, we introduce FinVerse, a finance-domain time-series forecasting benchmark that takes a first step toward more realistic evaluation. The released FinVerse data artifact contains 116,897 financial time series with 171.1M observations, of which 60,232 series with 17.4M observations are selected as evaluated targets based on their economic relevance to financial decisions. Unlike generic forecasting benchmarks that primarily emphasize uniform point-forecast or probabilistic accuracy, FinVerse defines 11 metric families comprising 78 evaluation metrics and assigns the most appropriate evaluation metrics to each individual time series based on its underlying economic meaning. Our analysis of 43 public time-series forecasting foundation models shows that strong performance under generic forecasting criteria does not necessarily translate into useful financial forecasts. This finding highlights the need for domain-aware benchmarks that evaluate models under objectives closer to real-world decision making.
Tags
Links
- Source: https://arxiv.org/abs/2608.03259v1
- Canonical: https://arxiv.org/abs/2608.03259v1
Trouble viewing inline? Open PDF directly →
Full Text
77,240 characters extracted from source content.
Expand or collapse full text
plain FinVerse: Financial Time-Series Benchmark Toward A More Realistic Evaluation for Financial Time-series Forecasting Jaehoon Lee, Jun Seo, Seunghan Lee, Tae Yoon Lim, Dongwan Kang, Hwanil Choi, Minjae Kim, Sungdong Yoo, Junhyeok Kang, Sangjun Han, Soonyoung Lee, Wonbin Ahn LG AI Research Corresponding author. 1 Introduction The recent success of large language models has renewed interest in foundation models across a broad range of domains. This trend has naturally extended to time-series modeling, where a growing body of work aims to build general-purpose models that can forecast, represent, or reason over temporal observations from many sources [das2024decoder, liu2024timer, woo2024unified, goswami2024moment, liu2025timerxl, taga2025timepfn, ansari2025chronos2]. As these models become larger and more general, the field increasingly needs evaluation protocols that can distinguish models that are merely accurate under generic statistical criteria from models that are useful for realistic downstream decisions. Well-defined natural language benchmarks have played this role in the development of language models. [wang2018glue, wang2019superglue, hendrycks2021mmlu, liang2022helm] In natural language processing, benchmark suites provide shared tasks, standardized data splits, and interpretable metrics that make it possible to compare models and track progress over time. More importantly, strong performance on a well-designed benchmark is expected to correlate with practical capability on real tasks. A benchmark therefore acts not only as a leaderboard, but also as a compact definition of what the community currently considers meaningful progress. Time-series forecasting does not yet have an equally satisfactory benchmark standard. Existing benchmarks, including widely used suites such as GIFT-Eval [aksu2024gifteval], have provided valuable common ground for comparing forecasting models. However, they typically evaluate heterogeneous time series with a largely uniform set of error-based forecast accuracy metrics. This design raises a practical question: if a decision maker selects the best-performing model on such a benchmark, should they expect reasonable performance on the decisions that actually depend on the forecast? We argue that the answer is often no, because the meaning of a good forecast changes across domains and target series. For example, in stock forecasting, correctly anticipating the direction or relative ranking of future returns can be more relevant than minimizing price error. For macroeconomic indicators such as CPI, both the direction of change and the acceleration or deceleration relative to the previous period may matter. A benchmark that applies the same MASE- and CRPS-centered evaluation to all series cannot fully capture these domain-specific requirements. To address this gap, we propose TimeVerse (Time-series benchmark uniVerse), an initiative toward more realistic evaluation of time-series forecasting. In this paper, as its first domain-specific benchmark, we introduce FinVerse, which focuses on financial time-series forecasting.111We plan to extend this benchmark to additional domains, such as medicine and energy, where realistic evaluation likewise requires domain-specific metrics. FinVerse is designed around the principle that each time series should be evaluated according to the decision context implied by its economic meaning. Rather than treating financial forecasting as a homogeneous error minimization problem, the benchmark organizes financial time series around three evaluation aspects: point-wise forecast accuracy, cross-sectional ranking quality, and portfolio-level backtesting. Across these three evaluation aspects, we define 11 metric families comprising 78 evaluation metrics and assign the most appropriate metric to each individual time series based on its underlying economic meaning. This design aims to provide a more realistic and practically informative view of model performance in financial forecasting. FinVerse collects financial time series across a broad hierarchy of scopes and categories. The benchmark covers inter-country series such as foreign exchange rates, commodities, crypto assets, and global market indices; country-level series such as fixed income, macroeconomic indicators, and market indicators; and individual-instrument series such as equities, ETFs, and firm fundamentals. This structure allows the benchmark to evaluate models on diverse financial time series that require different notions of predictive usefulness. The released FinVerse data artifact contains 116,897 financial time series with 171.1M observations. Among them, we select 60,232 economically meaningful target series with 17.4M observations as the fixed evaluation set, manually choosing series whose semantics make them important for financial forecasting and decision making. From the remaining historical data before the 2020-01-01 evaluation start date, FinVerse constructs a training set with 66,134 time series and 100.3M observations for foundation model pretraining and fine-tuning. Using 2020-01-01 as the evaluation start date, the benchmark evaluates selected financial time series. Table 1 summarizes the coverage of the FinVerse data artifact. Through this benchmark, we observe that models that perform well under generic time-series benchmarks do not necessarily produce useful results for financial forecasting tasks. These findings highlight the need for financial-domain benchmarks that evaluate models under decision-oriented settings rather than relying solely on generic error-based forecasting accuracy. Table 1: Coverage of the full FinVerse data artifact by scope and category. D, BD, W, M, Q, and A denote daily, business-daily, weekly, monthly, quarterly, and annual frequencies, respectively. Scope Category Series Points Frequencies Inter-country Foreign Exchange 25 294.4K BD Commodity 59 261.9K BD/W/M Crypto 4,700 5.0M D Global Liquidity 39 4.5K Q Global Risk 6 18.1K D/BD/M Market Index 22 166.4K BD Country Fixed Income 75 380.9K BD/W/M Market Indicators 32 270.8K D/BD/W/M Macro Indicators 199 165.2K D/BD/W/M/Q/A Individual ETF 25,531 40.5M BD Equity 30,745 121.7M BD Fundamentals 55,442 2.2M Q News Sentiment 22 75.0K D The main contributions of this report are as follows: • We identify a key limitation of existing time-series benchmarks: their reliance on uniform error-based metrics can obscure whether a model is useful for realistic domain decisions. • We introduce FinVerse, a large-scale finance benchmark consisting of 116,897 financial time series (171.1M observations). FinVerse addresses the limitations of existing time-series benchmarks by replacing one-size-fits-all error-based evaluation with decision-oriented metrics tailored to the underlying economic meaning of each financial time series. • We provide empirical evidence that strong performance on existing generic error-based time-series benchmarks does not necessarily imply reasonable performance on practical financial forecasting objectives. 2 Related Work 2.1 Time-Series Foundation Models The foundation model paradigm has recently expanded from language and vision to time-series modeling. Traditional forecasting models are often trained for a specific dataset, domain, frequency, or forecasting horizon, which limits their ability to transfer across heterogeneous temporal data. In contrast, recent time-series foundation models aim to pretrain on large and diverse collections of time series and then support zero-shot or few-shot forecasting on unseen datasets. Representative forecasting-oriented models include TimesFM [das2024decoder], Chronos [ansari2024chronos, ansari2025chronos2], Moirai [woo2024unified], Timer [liu2024timer, liu2025timerxl], and TimePFN [taga2025timepfn], which are pretrained on large collections of heterogeneous time series. The dominant architectural paradigm among recent time-series foundation models is the decoder-only Transformer, which adapts autoregressive sequence modeling from language models to temporal prediction. Given a historical sequence, these models generate future observations in an autoregressive manner, enabling flexible forecasting across diverse domains without task-specific architectural modifications. Representative examples include TimesFM, Chronos, and Timer. Beyond forecasting-oriented models, general-purpose representation models such as MOMENT [goswami2024moment] learn transferable time-series representations that can be adapted to a broad range of downstream tasks, including forecasting, classification, anomaly detection, imputation, and representation learning. Together, these advances have substantially improved the generalization ability of time-series models and accelerated the development of general-purpose time-series foundation models. 2.2 Time-Series Benchmarks Several benchmark efforts have shaped modern time-series forecasting evaluation. TSLib-style long-sequence forecasting protocols popularized standardized comparisons across canonical datasets [zhou2021informer]. The Monash Time Series Forecasting Archive broadened the evaluation landscape by collecting diverse real-world forecasting datasets under a unified archive and metadata format [godahewa2021monash]. More recently, LOTSA, introduced together with Moirai, scaled pretraining and evaluation toward a large collection of heterogeneous time-series datasets, supporting the development of universal forecasting models [woo2024unified]. GIFT-Eval is a representative and widely used benchmark for general time-series forecasting model evaluation [aksu2024gifteval]. It provides a standardized evaluation suite over heterogeneous datasets and has become a practical reference point for comparing recent time-series foundation models. However, its general-purpose design necessarily abstracts away many domain-specific objectives. In financial forecasting, the practical value of a forecast often depends on direction, ranking, risk, or portfolio-level outcomes rather than only generic error-based point or probabilistic forecast accuracy. FinVerse is designed to complement such general benchmarks by evaluating models under finance-specific metrics and decision contexts. 2.3 Financial Forecasting Forecasting is a central task in finance because many downstream decisions depend on expectations about future market behavior, risk, and economic conditions. Recent work has studied financial forecasting from several complementary angles, including spatial-temporal modeling for stock time series, adaptation under distribution shift for stock trend forecasting, and volatility prediction from options-derived signals [yan2024doublepath, zhao2023doubleadapt, soroka2025dataefficient]. These studies aim to improve forecasting performance on financial time series, whose non-stationary dynamics make accurate prediction particularly challenging. Recent benchmark efforts have also begun to systematize financial time-series forecasting evaluation. FinTSB provides a dedicated benchmark for financial time-series forecasting [hu2025fintsb], but its evaluation scope is primarily centered on stock data, and its prediction target focuses on next-day prices, which limits diversity in both data coverage and evaluation periods. In contrast, FinVerse collects financial time series across multiple data categories, including equity prices, exchange rates, commodities, crypto assets, fixed-income indicators, market indicators, macroeconomic indicators, ETFs, and fundamentals. It further supports multiple data frequencies and evaluation horizons, aiming to provide a broader and more practically useful evaluation setting for financial forecasting models. 3 FinVerse FinVerse is built as a finance-domain benchmark universe rather than as a simple collection of heterogeneous financial time series. The benchmark first organizes diverse financial signals into a hierarchical taxonomy defined by Scope, Category, and Detail Category. For each Detail Category–Frequency combination, FinVerse assigns evaluation metrics that reflect the semantic meaning and economic role of the corresponding time series. The metric assignments for each Detail Category–Frequency combination are summarized in Tables LABEL:tab:appendix_metric_assignment_intercountry_foreign_exchange–LABEL:tab:appendix_metric_assignment_individual_fundamentals. Section 3.1 describes how the benchmark is structured under this hierarchy and reports evaluation data coverage statistics. Section 3.2 then explains the domain-aware metrics used to evaluate financial time series, and Section 3.3 describes the evaluation periods and forecast horizons used to construct forecasting tasks. Figure 1: Observation coverage of the selected FinVerse evaluation targets by category and frequency. 3.1 Benchmark Universe Construction FinVerse organizes financial time series along a scope–category–detail category hierarchy. The benchmark contains three scopes. The inter-country scope covers cross-country global series, including Foreign Exchange, Commodity, Crypto, Global Liquidity, Global Risk, and Market Index categories. The country scope covers country-level economic and market conditions, including Fixed Income, Macro Indicators, and Market Indicators. The individual scope covers individual tradable instruments, firm-level targets, and sector-level signals, including ETF, Equity, Fundamentals, and News Sentiment categories. The detailed category-level coverage for each category is reported in Appendix Table LABEL:tab:appendix_detail_coverage. In total, the released full data artifact contains 3 scopes, 13 categories, and 101 detail categories with 116,897 financial series and 171.1M observations. The released data include daily, business-daily, weekly, monthly, quarterly, and annual frequencies. Note that the country and individual scopes are centered on the U.S. market, the world’s largest financial market, whereas the inter-country scope includes cross-country and global financial series. Among all collected financial time series, we manually select 60,232 target series (17.4M observations) to form the fixed evaluation set based on their underlying economic meaning and relevance to financial forecasting and decision-making. The resulting evaluation set spans 3 scopes, 10 evaluated categories, 67 evaluated detail categories, and 75 evaluated detail-category–frequency units. Figure 1 summarizes the observation-level coverage of the selected evaluation targets. The largest portion of the evaluation set consists of business-daily observations from equity (10.2M observations) and ETF (5.5M observations) series, reflecting the large number of individual instruments represented in these categories. Although country- and inter-country-level series, which are primarily available at monthly, quarterly, and annual frequencies, contain fewer observations, they remain essential components of the benchmark because they capture economically important macroeconomic and market conditions. 3.2 Benchmark Evaluation Metrics FinVerse uses metric families that correspond to three evaluation views: point-wise forecast quality, cross-sectional ranking quality, and realized portfolio performance. For a single series, let yty_t denote the realized value at time t, let y^t+h y_t+h denote the forecast for horizon h, and let O be the set of valid forecast origins. 3.2.1 Point-wise Forecasting Hit Ratio. Hit ratio measures directional correctness. For a comparison window p and forecast horizon h, define the realized and predicted changes as ro(p,h)=yo+hyo+h−p−1,r^o(p,h)=y^o+hyo+h−p−1.r_o^(p,h)= y_o+hy_o+h-p-1, r_o^(p,h)= y_o+hy_o+h-p-1. (1) The hit ratio is HR(p,h)=1||∑o∈[sign(r^o(p,h))=sign(ro(p,h))].HR^(p,h)= 1|O| _o I\! [sign\! ( r_o^(p,h) )=sign\! (r_o^(p,h) ) ]. (2) For first- and second-order directional changes, define Δro(p,h)=ro(p,h)−ro−1(p,h),Δr^o(p,h)=r^o(p,h)−ro−1(p,h), r_o^(p,h)=r_o^(p,h)-r_o-1^(p,h), r_o^(p,h)= r_o^(p,h)-r_o-1^(p,h), (3) and Δ2ro(p,h)=ro(p,h)−2ro−1(p,h)+ro−2(p,h),Δ2r^o(p,h)=r^o(p,h)−2ro−1(p,h)+ro−2(p,h). ^2r_o^(p,h)=r_o^(p,h)-2r_o-1^(p,h)+r_o-2^(p,h), ^2 r_o^(p,h)= r_o^(p,h)-2r_o-1^(p,h)+r_o-2^(p,h). (4) The derivative hit ratios are then ΔHR(p,h)=1||∑o∈[sign(r^o(p,h))=sign(ro(p,h))][sign(Δr^o(p,h))=sign(Δro(p,h))], ^(p,h)= 1|O| _o I\! [sign\! ( r_o^(p,h) )=sign\! (r_o^(p,h) ) ]I\! [sign\! ( r_o^(p,h) )=sign\! ( r_o^(p,h) ) ], (5) Δ2HR(p,h)=1||∑o∈[sign(r^o(p,h))=sign(ro(p,h))][sign(Δr^o(p,h))=sign(Δro(p,h))][sign(Δ2r^o(p,h))=sign(Δ2ro(p,h))]. ^2HR^(p,h)= 1|O| _o I\! [sign\! ( r_o^(p,h) )=sign\! (r_o^(p,h) ) ]I\! [sign\! ( r_o^(p,h) )=sign\! ( r_o^(p,h) ) ]I\! [sign\! ( ^2 r_o^(p,h) )=sign\! ( ^2r_o^(p,h) ) ]. (6) The comparison window p specifies the reference interval used to compute the realized and predicted changes (e.g., week-over-week, month-over-month, quarter-over-quarter, half-year-over-half-year, or year-over-year), whereas the forecast horizon h specifies how far into the future the prediction is evaluated. For example, p=1yp=1y computes the year-over-year change relative to the observation one year earlier, while h=1yh=1y evaluates the prediction one year ahead. Different financial series use different (p,h)(p,h) combinations according to their sampling frequency and economic interpretation. The complete (p,h)(p,h) assignments for each financial series and evaluation metric are provided in Tables LABEL:tab:appendix_metric_assignment_intercountry_foreign_exchange–LABEL:tab:appendix_metric_assignment_individual_fundamentals. HR is motivated by the fact that many financial decisions depend more on predicting the direction of movement than on reconstructing the exact future value. For example, in asset-price forecasting, correctly predicting whether the price will rise or fall after a given horizon is often more important than minimizing the level error. The proposed ΔHR and Δ2HR ^2HR extend this directional criterion to first- and second-order changes, respectively. These variants are particularly useful for macroeconomic indicators such as CPI, where long-term trends make it important to predict not only whether the year-over-year value will increase or decrease, but also whether the rate of change is accelerating or decelerating relative to the previous period. Mean Absolute Scaled Error. MASE is included because scaled error metrics have been widely adopted in time-series forecasting benchmarks and provide a scale-independent measure that is comparable across heterogeneous series. Unlike the original MASE, which scales forecast errors using the in-sample error of a seasonal or naive forecast, FinVerse adapts the scaling baseline to better reflect financial forecasting objectives, where the magnitude of future values is often as important as their direction. While the proposed metrics share the same intuition as the original MASE—measuring forecast error relative to a simple baseline predictor—the baseline is defined directly for the prediction target under evaluation. This design is motivated by financial time series (e.g., price), which often exhibit weak or unstable seasonality compared with many benchmark datasets, making persistence-based baselines more appropriate than the seasonal-naive scaling used in the original MASE. Specifically, three variants are considered. MASErelMASE_rel measures forecast error on relative changes against a no-change relative baseline. Since ro(p,h)r_o^(p,h) denotes the relative change from yo+h−py_o+h-p to yo+hy_o+h, this baseline corresponds to predicting zero relative change (i.e., a gross return of one). FinVerse computes MASErel(p,h)=1||∑o∈|r^o(p,h)−ro(p,h)|1||∑o∈|ro(p,h)|.MASE_rel^(p,h)= 1|O| _o | r_o^(p,h)-r_o^(p,h) | 1|O| _o |r_o^(p,h) |. (7) MASEabsMASE_abs evaluates raw-value forecasts against a no-change raw-value baseline defined over the same comparison window. In other words, the baseline predicts the future value by repeating the observation at yo+h−py_o+h-p: MASEabs(p,h)=1||∑o∈|y^o+h−yo+h|1||∑o∈|yo+h−yo+h−p|.MASE_abs^(p,h)= 1|O| _o | y_o+h-y_o+h | 1|O| _o |y_o+h-y_o+h-p |. (8) Finally, MASEoriMASE_ori evaluates raw-value forecasts against the conventional origin-value persistence baseline, where the most recently observed value is repeated for all future horizons. Since the raw-value metric always uses p=hp=h, the baseline reduces to persistence from the forecast origin: MASEori=1||H∑o∈∑h=1H|y^o+h−yo+h|1||H∑o∈∑h=1H|yo−yo+h|.MASE_ori= 1|O|H _o _h=1^H | y_o+h-y_o+h | 1|O|H _o _h=1^H |y_o-y_o+h |. (9) We include MASE-based metrics because, although directional correctness is often the primary objective in financial forecasting, predictions with unrealistic magnitudes are of limited practical value. For example, in foreign exchange forecasting, where price movements are typically small, a model that consistently predicts excessively large changes may achieve reasonable directional accuracy while remaining unsuitable for real-world use. Accordingly, point-wise evaluation should consider both directional behavior and magnitude accuracy. These three MASE variants capture complementary aspects of financial forecasting. MASErelMASE_rel evaluates the accuracy of predicting relative changes (e.g., returns), whereas MASEabsMASE_abs measures the accuracy of predicting the raw values themselves. Furthermore, MASErelMASE_rel and MASEabsMASE_abs are evaluated only at financially meaningful forecasting horizons (e.g., one week, one month, one quarter, half a year, and one year ahead), while MASEoriMASE_ori measures forecasting error over the entire prediction horizon. 3.2.2 Cross-sectional Ranking Information Coefficient. The information coefficient evaluates cross-sectional ranking quality. For a set of assets oS_o available at origin o, FinVerse computes a Spearman rank correlation between predicted and realized future changes: ICo(p,h)=ρSpearman(r^i,o(p,h)i∈o,ri,o(p,h)i∈o).IC_o^(p,h)= _Spearman ( \ r_i,o^(p,h) \_i _o, \r_i,o^(p,h) \_i _o ). (10) IC is selected because portfolio and allocation decisions often depend on relative ordering across assets rather than independent point accuracy for each series. This is used for groups where relative ranking is meaningful, such as currencies, commodities, crypto assets, global market indices, ETFs, equities, selected fixed-income groups, and selected fundamentals. 3.2.3 Real-world Portfolio Backtesting Portfolio evaluation converts forecasts into an equal-weight long-only strategy. At each rebalance origin, the benchmark ranks assets by the horizon-h predicted signal and selects the top 10% fraction o(h)⊂oP_o^(h) _o. For a holding horizon h, the selected set o(h)P_o^(h) is reconstructed at rebalance origins spaced h frequency steps apart and held for the next h steps. To avoid depending on a single calendar start date, FinVerse runs this backtest for every possible offset within the h-step window and averages the resulting portfolio statistics. Thus, for a one-month horizon, starts such as January 1 and January 2 define separate monthly-rebalanced paths rather than discarding all but one start date. The realized portfolio return for a rebalance origin is Ro(h)=1|o(h)|∑i∈o(h)ri,o(h,h).R_o^(h)= 1|P_o^(h)| _i _o^(h)r_i,o^(h,h). (11) Let h=0,…,h−1A_h=\0,…,h-1\ denote the offset set, and let os,1,…,os,Kso_s,1,…,o_s,K_s denote the rebalance origins in chronological order for offset s, where KsK_s is the number of valid rebalancing periods in that offset path. The cumulative net asset value for offset s is NAVs,k(h)=∏j=1k(1+Ros,j(h)).NAV_s,k^(h)= _j=1^k (1+R_o_s,j^(h) ). (12) The resulting return sequence is summarized with annualized return, annualized volatility, Sharpe ratio, and maximum drawdown for each offset, and the reported metric is the average across offsets: Metric(h)=1|h|∑s∈hMetrics(h).Metric^(h)= 1|A_h| _s _hMetric_s^(h). (13) Here, h denotes the holding horizon. The annualization factor is selected directly from the portfolio holding horizon: Ah∈52,12,4,2,1A_h∈\52,12,4,2,1\ (14) for one-week, one-month, one-quarter, half-year, and one-year horizons, respectively. For each offset s, the annualized return, annualized volatility, Sharpe ratio, and maximum drawdown are computed as AnnualReturns(h)=(∏j=1Ks(1+Ros,j(h)))AhKs−1,AnnualVolatilitys(h)=Std(Ros,j(h))Ah,AnnualReturn_s^(h)= ( _j=1^K_s (1+R_o_s,j^(h) ) ) A_hK_s-1, _s^(h)=Std\! (R_o_s,j^(h) ) A_h, (15) Sharpes(h)=[Ros,j(h)]Std(Ros,j(h))Ah,MDDs(h)=mink(NAVs,k(h)maxℓ≤kNAVs,ℓ(h)−1).Sharpe_s^(h)= E[R_o_s,j^(h)]Std(R_o_s,j^(h)) A_h, _s^(h)= _k ( NAV_s,k^(h) _ ≤ kNAV_s, ^(h)-1 ). (16) These metrics are included because financial forecasts are ultimately useful only when they can support stable downstream decisions under realistic return and risk tradeoffs. We acknowledge that many investment strategies are possible and that benchmark outcomes can vary with the chosen strategy; FinVerse therefore starts with a simple portfolio rule and leaves richer strategy families for future extensions. Overall, the benchmark defines six metric families for point-wise forecasting, one metric family for cross-sectional ranking, and four metric families for realized portfolio evaluation. Combined with the predefined (p,h)(p,h) or (h)(h) combinations, these yield a total of 78 evaluation metrics. 3.3 Evaluation Periods and Forecast Horizons The default training artifact contains observations before the 2020-01-01 evaluation start date. As summarized in Table 2, we define the prediction length and the minimum lookback window size for each sampling frequency. The prediction lengths are chosen to correspond to financially meaningful forecasting horizons (i.e., one week, one month, one quarter, half a year, and one year ahead), while the minimum lookback window sizes are selected to provide sufficient historical context for making meaningful forecasts at the corresponding horizons. Table 2: Frequency-specific prediction lengths and minimum lookback requirements. Frequency Prediction Length Minimum Lookback Daily 365 365 Business-daily 260 260 Weekly 52 52 Monthly 12 12 Quarterly 4 4 Annual 1 4 4 Evaluation with FinVerse 4.1 Foundation Models We evaluate FinVerse with a broad set of recent time-series foundation models and forecasting checkpoints. The evaluated models include TimesFM 1.0 200M, TimesFM 2.0 500M, and TimesFM 2.5 200M [das2024decoder]; Chronos-2, Chronos-2 Synth, Chronos-Bolt Small, and Chronos-Bolt Base [ansari2024chronos, ansari2025chronos2]; Moirai 1.1-R Small, Moirai 1.1-R Base, Moirai 1.1-R Large, and Moirai 2.0-R Small [woo2024unified]; Timer-S1 [liu2024timer, liu2025timerxl]; TempoPFN [hollmann2025tempopfn]; Toto Open Base 1.0, Toto 2.0 4M, Toto 2.0 22M, Toto 2.0 313M, Toto 2.0 1B, and Toto 2.0 2.5B [toto2025]; TiRex 1.1, TiRex 2 Pretrain, and TiRex 2 ZS [dooley2025tirex]; VisionTS [chen2024visionts]; Sundial Base 128M [liu2025sundial]; YingLong 6M, YingLong 50M, YingLong 110M, and YingLong 300M [li2025yinglong]; Kairos-10M, Kairos-23M, and Kairos-50M [kairos2025]; Reverso Small [reverso2025]; TTM R1, TTM R2, and TTM R3 [ekambaram2024tinytimemixer]; PatchTST-FM R1 and Granite PatchTST-FM R1 [nie2023patchtst]; FlowState and Granite FlowState R1 [ibm2025flowstate]; PatchFM; T0 Alpha; Super Linear; and CleanTS-65M. 4.2 Overall Rank Aggregation FinVerse contains heterogeneous evaluation metrics with different numerical characteristics, and the numbers of time series vary substantially across detail category–frequency combinations. Directly averaging raw metric values would therefore overemphasize certain metrics and categories. To address this issue, FinVerse adopts a rank-based aggregation strategy rather than directly averaging raw metric values. Specifically, we first compute the average score of each evaluation metric within every detail category–frequency combination. Models are then ranked separately for each metric within each detail category–frequency combination. Next, for each of the three evaluation groups—Point-wise Forecasting, Cross-sectional Ranking, and Realized Portfolio Backtesting—we compute the geometric mean of the corresponding metric ranks, denoted by K. The models are then ranked according to the K value of each evaluation group. Finally, the overall rank is obtained by summing the three group-level ranks. This aggregation strategy enables heterogeneous metrics with different scales and properties to contribute more evenly to the final evaluation while preventing categories containing many time series from dominating the benchmark. As a result, smaller but economically important categories are appropriately reflected in the overall ranking. 4.3 Performance in FinVerse Figure 2: Relationship between model size and overall FinVerse rank. Lower rank indicates better performance. 4.3.1 Overall Benchmark Performance Figure 2 compares model size with the overall FinVerse rank computed from the three aspect placement ranks. Reverso Small achieves the best overall rank, followed by Chronos-2 Synth, TiRex 1.1, TimesFM 2.5 200M, and Chronos-2. In contrast, larger checkpoints do not consistently dominate: several Toto, Timer, and YingLong variants remain behind smaller or medium-sized models in the overall ranking. These results suggest that financial decision-oriented evaluation cannot be explained by parameter count alone. Instead, model pretraining data, inductive biases, horizon handling, and robustness to financial distribution shifts appear to contribute at least as much as model scale. Figure 3: Comparison between FinVerse and GIFT-Eval rankings for matched models. The GIFT-Eval rank is the average of the public GIFT-Eval MASE rank and probabilistic forecast rank. The star marks the best-ranked model in FinVerse, and the X marker marks the best-ranked model in GIFT-Eval. Figure 3 compares the overall FinVerse rank with the corresponding GIFT-Eval ranking for models evaluated on both benchmarks. Across the 43 matched models, the two rankings have only a moderate correlation (Pearson r=0.40r=0.40), indicating that they are related but far from interchangeable. For example, Chronos-2, TiRex 1.1, and TimesFM 2.5 200M rank strongly on both benchmarks, while Reverso Small is the strongest FinVerse model despite a weaker GIFT-Eval rank. These differences indicate that strong performance on a generic time-series benchmark does not necessarily translate into superior performance under FinVerse’s decision-oriented financial evaluation. 4.3.2 Performance Across Evaluation Aspects Figure 4: Top-20 models ranked separately under point-wise, cross-sectional, and portfolio-oriented evaluation aspects. Figure 4 reports model rankings separately for the three evaluation aspects: point-wise forecasting, cross-sectional ranking, and portfolio backtesting. TiRex 1.1 performs best in the point-wise aspect, followed by TiRex 2 Pretrain, TiRex 2 ZS, PatchFM, and Reverso Small. Cross-sectional ranking favors Reverso Small, Chronos-2 Synth, TiRex 1.1, TimesFM 2.5 200M, and TiRex 2 Pretrain. Portfolio evaluation again favors a different set of models: Chronos-2 Synth ranks first, followed by TimesFM 2.0 500M, T0 Alpha, Reverso Small, and Moirai 1.1-R Large. These results demonstrate that strong point-wise forecasting performance does not necessarily imply superior cross-sectional ranking or portfolio performance. Figure 5: Pairwise relationship between point-wise, cross-sectional, and portfolio ranks at the model–detail-category–frequency level. Figure 5 further examines the relationships among the three evaluation aspects at the model–detail-category–frequency level. Point-wise and cross-sectional rankings are strongly related in this view (r=0.72r=0.72). The point-wise–portfolio relationship is weaker (r=0.37r=0.37), while the cross-sectional–portfolio relationship is moderately aligned (r=0.60r=0.60). Portfolio-level outcomes therefore capture information beyond conventional per-series forecasting accuracy and cross-sectional rank correlation. Overall, the three evaluation aspects are complementary rather than redundant, supporting the need for a benchmark that evaluates forecasting models from multiple financial perspectives. Table 3: Metric-wise model performance across point-wise, cross-sectional, and portfolio-oriented metrics using averaged raw metric values. Best performance in each metric column is bolded. Model HR Δ Δ2 ^2HR MASErelMASE_rel MASEabsMASE_abs MASEoriMASE_ori IC Return Sharpe MDD Vol. Reverso Small 0.673 0.542 0.481 11.548 11.652 1.573 0.146 0.352 0.638 -0.282 1.173 Chronos-2 Synth 0.646 0.550 0.495 17.969 7.936 1.612 0.153 0.946 0.632 -0.227 2.786 TiRex 1.1 0.666 0.577 0.518 13.713 7.162 1.548 0.152 0.390 0.603 -0.297 1.845 TimesFM 2.5 200M 0.653 0.568 0.509 12.242 10.373 1.596 0.147 0.317 0.605 -0.298 1.222 Chronos-2 0.659 0.560 0.495 21.008 7.474 1.505 0.072 0.657 0.586 -0.286 2.622 Chronos-Bolt Base 0.645 0.559 0.493 19.900 6.783 1.576 0.100 0.353 0.637 -0.303 1.172 Chronos-Bolt Small 0.644 0.551 0.491 19.643 6.898 1.561 0.099 0.592 0.970 -0.307 2.415 T0 Alpha 0.621 0.566 0.507 15.088 10.151 1.838 0.129 0.800 0.588 -0.285 2.976 TiRex 2 Pretrain 0.652 0.574 0.515 16.253 6.454 1.547 0.151 0.488 0.553 -0.300 2.346 Toto 2.0 2.5B 0.661 0.547 0.483 15.239 1.524 1.809 0.106 0.539 0.638 -0.296 2.166 TiRex 2 ZS 0.650 0.572 0.513 17.120 6.481 1.533 0.150 0.441 0.567 -0.306 2.079 VisionTS 0.639 0.568 0.506 19.033 6.459 1.562 0.108 0.619 0.646 -0.321 2.543 PatchFM 0.672 0.565 0.500 4.105 11.809 1.653 0.127 0.423 0.606 -0.303 1.837 Toto 2.0 4M 0.662 0.556 0.493 14.976 4.521 1.661 0.100 0.602 0.619 -0.285 2.465 TimesFM 2.0 500M 0.534 0.417 0.365 211.602 163.602 205.471 0.131 0.297 0.698 -0.266 1.037 Toto 2.0 1B 0.661 0.553 0.489 12.339 3.117 1.819 0.104 0.471 0.609 -0.298 2.167 Kairos-50M 0.659 0.548 0.481 28.325 20.003 1.730 0.104 0.560 0.605 -0.289 2.309 TempoPFN 0.668 0.514 0.446 35.542 15.494 1.648 0.080 0.573 0.575 -0.302 2.757 Granite FlowState R1 0.642 0.563 0.486 55.796 4.328 2.218 0.048 0.611 0.645 -0.305 2.686 TimesFM 1.0 200M 0.622 0.556 0.499 19.547 10.575 7.237 0.134 0.634 0.917 -0.299 2.122 Super Linear 0.592 0.534 0.474 15.109 6.970 1.674 0.119 0.931 0.585 -0.318 3.042 Kairos-10M 0.655 0.506 0.441 61.964 23.403 2.114 0.077 0.840 0.624 -0.295 2.695 Toto 2.0 22M 0.658 0.554 0.488 11.000 3.499 1.676 0.104 0.553 0.594 -0.288 2.340 FlowState 0.641 0.550 0.476 47.539 5.101 2.080 0.025 0.748 0.619 -0.302 2.864 Toto 2.0 313M 0.660 0.557 0.490 15.043 5.001 1.856 0.096 0.490 0.597 -0.297 2.066 Timer-S1 0.606 0.533 0.477 16.586 10.489 1.785 0.132 0.513 2.633 -0.314 2.021 Toto Open Base 1.0 0.596 0.495 0.434 15.642 5.919 1.845 0.091 0.728 0.570 -0.293 2.909 Moirai 1.1-R Large 0.532 0.438 0.388 48.699 10.707 2.238 0.108 0.355 0.597 -0.278 1.127 Moirai 1.1-R Base 0.538 0.440 0.385 24.502 6.860 1.957 0.101 0.366 2.667 -0.288 1.255 Moirai 1.1-R Small 0.548 0.445 0.393 34.274 11.824 1.840 0.117 0.357 2.681 -0.300 1.208 PatchTST-FM R1 0.565 0.440 0.383 20.755 12.921 2.630 0.107 0.900 2.637 -0.304 2.997 Sundial Base 128M 0.603 0.540 0.483 16.902 9.571 1.832 0.123 0.734 0.561 -0.318 2.507 Kairos-23M 0.642 0.522 0.457 39.604 8.560 1.857 0.079 0.810 0.570 -0.305 2.878 Granite PatchTST-FM R1 0.539 0.433 0.377 19.615 12.962 2.463 0.109 0.756 0.537 -0.305 2.987 Moirai 2.0-R Small 0.569 0.494 0.442 9.469 6.866 2.043 0.093 0.568 0.544 -0.294 2.583 TTM R3 0.531 0.427 0.380 5.792 7.628 2.048 0.121 0.539 0.579 -0.322 2.277 TTM R1 0.586 0.450 0.384 31.108 8.051 7.918 0.085 0.790 0.582 -0.313 2.749 CleanTS-65M 0.567 0.477 0.413 19.269 11.818 2.159 0.098 0.763 0.543 -0.313 2.944 TTM R2 0.611 0.459 0.393 30.052 7.214 2.054 0.103 0.583 2.662 -0.329 2.069 YingLong 50M 0.406 0.276 0.240 20.777 12.918 2.627 0.054 0.691 0.560 -0.313 2.589 YingLong 300M 0.404 0.273 0.237 20.621 12.927 2.624 0.052 0.786 0.549 -0.310 2.860 YingLong 110M 0.405 0.274 0.239 21.437 12.920 2.618 0.047 0.945 0.546 -0.313 2.874 YingLong 6M 0.406 0.275 0.239 20.852 12.897 2.617 0.051 0.670 0.550 -0.315 2.643 4.3.3 Metric-wise Analysis Table 3 further decomposes model performance by metric family using averaged raw metric values. Even within the point-wise evaluation aspect, the best-performing model differs across metrics. Reverso Small achieves the strongest HR value, TiRex 1.1 leads Δ and Δ2 ^2HR, PatchFM leads MASErelMASE_rel, Toto 2.0 2.5B leads MASEabsMASE_abs, and Chronos-2 leads MASEoriMASE_ori. For cross-sectional and portfolio-oriented metrics, Chronos-2 Synth achieves the strongest IC, Return, and MDD values, Moirai 1.1-R Small leads Sharpe, and TimesFM 2.0 500M achieves the lowest Volatility value. These results further demonstrate that models preferred by point-wise metrics are not necessarily those preferred by ranking-based or portfolio-oriented evaluation. Overall, the metric-wise analysis reinforces the central motivation of FinVerse. Point-wise forecasting accuracy, directional correctness, cross-sectional ranking quality, and portfolio performance each capture distinct aspects of financial forecasting utility, and no single model consistently dominates across all evaluation criteria. 5 Known Limitations and Future Work FinVerse is designed as an extensible benchmark rather than a closed final artifact. This release prioritizes a broad, reproducible evaluation interface, but several limitations remain in temporal alignment and domain coverage. 5.1 Calendar and Release-Date Alignment FinVerse currently normalizes weekly, monthly, quarterly, and annual observations to standard calendar period ends. This convention provides a consistent origin grid across heterogeneous series, but it does not fully preserve series-specific reporting calendars. For example, U.S. fiscal-year series may end in September rather than December, USDA crop-year series may follow agricultural calendars ending in June, and weekly series may be released on weekdays other than the normalized weekly anchor. A related limitation is point-in-time availability. Some economic indicators are assigned to the period they measure, even though the value is released only later. For example, a January CPI observation is economically associated with January, but it is typically released in early to mid-February. If the observation is timestamped at January 31 without an explicit release-date field, the benchmark may treat information as available earlier than it would have been in a real forecasting setting. 5.2 Future Extensions Several extensions are planned. First, FinVerse should improve point-in-time temporal alignment by adding calendar-aware preprocessing and release-date metadata for fiscal, crop-year, macroeconomic, and filing-date-sensitive series. Second, FinVerse should broaden coverage for index constituents, sector indices, survey-based indicators, fixed-income instruments, and country-specific datasets beyond the United States. Third, future releases should expand the set of decision-oriented evaluation tasks and portfolio construction protocols so that the benchmark can cover a wider range of practical financial use cases. Beyond finance, FinVerse is intended to be the first domain-specific member of a broader time-series benchmark universe. The same principle can be applied to domains such as medicine and energy, where realistic evaluation also requires metrics aligned with domain-specific decisions rather than a single generic point-forecast score. 6 Conclusion This report introduced FinVerse, a finance-domain benchmark in the broader Time-Series Benchmark Universe. The benchmark is motivated by a gap between generic error-based time-series forecasting accuracy and the decision-oriented objectives that often determine usefulness in financial settings. Instead of evaluating all series through a single point-forecast metric family, FinVerse organizes financial forecasting around three complementary aspects: per-series accuracy, cross-sectional ranking, and portfolio backtesting. FinVerse is built from released data artifacts that share a common atomic time-series schema. The full data artifact contains 116,897 financial time series with 171.1M observations, including 60,232 evaluated target series with 17.4M observations that are selected according to their economic meaning and relevance to financial decisions. Using 2020-01-01 as the evaluation start date, the benchmark evaluates 75 detail-category–frequency units using 11 metric families expanded into 78 metric columns. For model development, the pre-evaluation portion before 2020-01-01 provides 66,134 training time series with 100.3M observations. This structure allows heterogeneous financial data to be represented through a common atomic universe while assigning metrics according to the economic meaning of each series. Our empirical results over 43 public forecasting models show that model quality changes substantially depending on the financial objective being measured. Reverso Small achieves the best overall FinVerse rank, Chronos-2 Synth leads the portfolio aspect, and TiRex 1.1 leads the point-wise aspect. Point-wise accuracy, cross-sectional ranking quality, and portfolio outcomes are correlated but not fully aligned, and no single model dominates all metric families. These findings support the central premise of FinVerse: strong error-based forecasting performance alone does not necessarily imply usefulness for financial decisions involving direction, relative ranking, risk, or realized portfolio returns. References Appendix A Detailed Data Coverage This appendix provides the detailed coverage table for the released FinVerse data. Table LABEL:tab:appendix_detail_coverage reports, for each scope, category, detail category, and frequency, whether the group contains evaluated target series, together with the number of series, observed data points in the full data artifact, and valid evaluation points after forecast-origin filtering. Table 4: Detailed full-data coverage by scope, category, detail category, frequency, evaluation-target status, observed points, and valid evaluation points. Scope Category Detail Category Frequency Eval. Series Points Eval. Points Inter-country Foreign Exchange FX Index Business-daily × 4 30,316 0 Inter-country Foreign Exchange FX Pair Business-daily ✓ 21 264,091 34,503 Inter-country Commodity Commodity ETF Broad Business-daily ✓ 2 10,433 3,298 Inter-country Commodity Commodity ETF Single Business-daily ✓ 5 25,154 8,245 Inter-country Commodity Commodity ETF Thematic Business-daily ✓ 2 8,149 3,298 Inter-country Commodity Future Front Business-daily ✓ 27 173,716 43,601 Inter-country Commodity Inventory Stock Weekly ✓ 7 10,985 2,303 Inter-country Commodity Spot Business-daily ✓ 3 28,307 4,932 Inter-country Commodity Spot Monthly ✓ 13 5,123 975 Inter-country Crypto Crypto Pair Close (Other) Daily × 831 794,412 0 Inter-country Crypto Crypto Pair Close (Top 100) Daily ✓ 100 179,443 120,963 Inter-country Crypto Crypto Pair High Daily × 932 959,788 0 Inter-country Crypto Crypto Pair Low Daily × 931 959,424 0 Inter-country Crypto Crypto Pair Open Daily × 931 968,233 0 Inter-country Crypto Crypto Pair TVL Daily × 24 33,254 0 Inter-country Crypto Crypto Pair Volume Daily × 951 1,082,920 0 Inter-country Global Liquidity FX Year Over Year Change Quarterly × 6 801 0 Inter-country Global Liquidity Stock Quarterly × 15 1,692 0 Inter-country Global Liquidity Year Over Year Change Quarterly × 18 1,965 0 Inter-country Global Risk Credit Spread Business-daily × 1 781 0 Inter-country Global Risk Leading Indicator Monthly × 2 1,489 0 Inter-country Global Risk Uncertainty Daily × 1 15,092 0 Inter-country Global Risk Uncertainty Monthly × 2 694 0 Inter-country Market Index Country Index Business-daily ✓ 20 155,667 31,391 Inter-country Market Index Regional Index Business-daily ✓ 2 10,724 3,304 Country Fixed Income Bond Total Return Index Business-daily ✓ 10 7,830 5,230 Country Fixed Income Breakeven Business-daily ✓ 2 12,172 3,304 Country Fixed Income Breakeven Monthly ✓ 2 455 150 Country Fixed Income Corporate Yield Business-daily ✓ 2 21,823 3,302 Country Fixed Income Corporate Yield Monthly ✓ 2 2,574 150 Country Fixed Income Credit Spread Business-daily ✓ 10 7,830 5,230 Country Fixed Income Forward Inflation Business-daily ✓ 1 6,086 1,652 Country Fixed Income Inflation Expectation Monthly ✓ 7 3,949 530 Country Fixed Income Money Market Business-daily ✓ 8 20,958 12,543 Country Fixed Income Mortgage Rate Weekly ✓ 3 5,617 812 Country Fixed Income Swap Rate Business-daily × 7 29,820 0 Country Fixed Income TIPS Yield Business-daily ✓ 4 22,478 6,604 Country Fixed Income Term Spread Business-daily ✓ 2 24,587 3,304 Country Fixed Income Treasury Bill Business-daily × 4 60,333 0 Country Fixed Income Treasury Yield Business-daily ✓ 11 154,357 18,161 Country Market Indicators Fin Conditions Weekly × 6 17,316 0 Country Market Indicators Fin Stress Weekly × 1 1,687 0 Country Market Indicators Fin Stress Monthly × 1 434 0 Country Market Indicators Implied Volatility Business-daily ✓ 9 55,495 13,361 Country Market Indicators Leading Index Monthly × 1 794 0 Country Market Indicators Money Spread Business-daily × 1 9,407 0 Country Market Indicators Recession Flag Daily × 1 62,606 0 Country Market Indicators Recession Flag Monthly × 1 2,056 0 Country Market Indicators Recession Prob Monthly × 1 705 0 Country Market Indicators Sentiment Business-daily × 1 9,478 0 Country Market Indicators US Equity Index Business-daily ✓ 9 110,839 14,868 Country Macro Indicators Activity Monthly ✓ 10 5,621 744 Country Macro Indicators Banking Weekly × 2 5,562 0 Country Macro Indicators Banking Monthly ✓ 2 1,902 150 Country Macro Indicators Capacity Utilization By Industry Monthly ✓ 11 8,385 825 Country Macro Indicators Credit Monthly ✓ 2 1,696 148 Country Macro Indicators Current Account Quarterly ✓ 4 650 48 Country Macro Indicators Energy Activity Monthly ✓ 6 2,557 397 Country Macro Indicators Energy Refinery Weekly × 5 7,302 0 Country Macro Indicators External Monthly ✓ 6 2,460 444 Country Macro Indicators External Quarterly ✓ 3 951 75 Country Macro Indicators FDI Flows Quarterly ✓ 1 317 24 Country Macro Indicators FDI Flows Annual ✓ 1 80 4 Country Macro Indicators Financial Flows Quarterly ✓ 1 316 24 Country Macro Indicators Fiscal Monthly ✓ 1 546 75 Country Macro Indicators Fiscal Quarterly ✓ 1 240 24 Country Macro Indicators Fiscal Annual ✓ 2 222 8 Country Macro Indicators Freight Monthly ✓ 1 313 73 Country Macro Indicators Gdp Quarterly ✓ 2 634 50 Country Macro Indicators Housing Monthly ✓ 4 2,367 219 Country Macro Indicators Housing Price Monthly ✓ 10 3,099 747 Country Macro Indicators Housing Price Quarterly × 1 204 0 Country Macro Indicators Income Monthly ✓ 3 2,421 225 Country Macro Indicators Industrial Production By Industry Monthly ✓ 28 19,776 2,100 Country Macro Indicators Inflation Monthly ✓ 10 6,793 675 Country Macro Indicators International Investment Position Quarterly ✓ 5 400 120 Country Macro Indicators International Investment Position Annual ✓ 1 50 4 Country Macro Indicators Inventory Monthly ✓ 7 2,870 518 Country Macro Indicators Labor Weekly × 2 6,189 0 Country Macro Indicators Labor Monthly ✓ 14 7,872 1,046 Country Macro Indicators Labor By Industry Monthly ✓ 16 15,744 1,200 Country Macro Indicators Money Weekly × 1 1,220 0 Country Macro Indicators Money Monthly ✓ 5 4,035 375 Country Macro Indicators Policy Rate Business-daily × 1 18,453 0 Country Macro Indicators Policy Rate Daily × 3 22,267 0 Country Macro Indicators Portfolio Flows Quarterly ✓ 7 2,231 168 Country Macro Indicators Productivity Quarterly ✓ 3 948 72 Country Macro Indicators Regional Fed Monthly ✓ 2 994 152 Country Macro Indicators Saving Monthly ✓ 1 807 75 Country Macro Indicators Sentiment Monthly ✓ 3 1,763 200 Country Macro Indicators Spending Monthly ✓ 6 3,690 450 Country Macro Indicators Stress Monthly ✓ 1 434 75 Country Macro Indicators TIC Flows Quarterly ✓ 4 851 95 Individual ETF ETF OHLCV Close Business-daily ✓ 5,107 8,100,873 3,919,924 Individual ETF ETF OHLCV High Business-daily × 5,107 8,098,720 0 Individual ETF ETF OHLCV Low Business-daily × 5,107 8,095,707 0 Individual ETF ETF OHLCV Open Business-daily × 5,107 8,108,763 0 Individual ETF ETF OHLCV Volume Business-daily × 5,103 8,143,493 0 Individual Equity Equity OHLCV Close Business-daily ✓ 6,143 24,193,386 7,743,280 Individual Equity Equity OHLCV High Business-daily × 6,144 24,189,072 0 Individual Equity Equity OHLCV Low Business-daily × 6,144 24,106,760 0 Individual Equity Equity OHLCV Open Business-daily × 6,145 24,253,740 0 Individual Equity Equity OHLCV Volume Business-daily × 6,169 24,986,230 0 Individual Fundamentals Capital Expenditures Quarterly ✓ 3,442 140,619 60,843 Individual Fundamentals Cash and Cash Equivalents Quarterly ✓ 4,642 206,110 84,757 Individual Fundamentals EPS Basic Quarterly ✓ 4,185 169,330 75,581 Individual Fundamentals EPS Diluted Quarterly ✓ 4,166 166,806 74,206 Individual Fundamentals Long-Term Debt Quarterly ✓ 3,249 100,936 43,229 Individual Fundamentals Net Income Quarterly ✓ 4,357 182,999 80,909 Individual Fundamentals Operating Cash Flow Quarterly ✓ 4,359 185,546 83,997 Individual Fundamentals Operating Income Quarterly ✓ 3,707 151,714 67,678 Individual Fundamentals R And D Expense Quarterly × 1,851 64,374 0 Individual Fundamentals Revenue Quarterly ✓ 3,677 99,475 60,974 Individual Fundamentals Shares Diluted Quarterly × 4,173 168,275 0 Individual Fundamentals Stockholders Equity Quarterly ✓ 4,697 210,564 93,125 Individual Fundamentals Total Assets Quarterly ✓ 4,699 196,491 91,002 Individual Fundamentals Total Liabilities Quarterly ✓ 4,238 160,730 77,746 Individual News Sentiment Sentiment Tone Daily × 11 37,488 0 Individual News Sentiment Sentiment Volume Daily × 11 37,488 0 Appendix B Metric Assignment by Data Category Appendix B expands the metric assignment used by FinVerse. Table 5 maps each compact variant ID to the corresponding (p,h)(p,h) combinations. Tables LABEL:tab:appendix_metric_assignment_intercountry_foreign_exchange–LABEL:tab:appendix_metric_assignment_individual_fundamentals then provide the detailed metric assignment for each evaluated scope-category group. Each table lists the detail category, frequency, number of evaluation points, evaluation aspect, metric families, and (p,h)(p,h) variant used for that task. Together, these tables specify how FinVerse connects each financial data group to the metrics that are economically meaningful for that group. Interpretation of (p,h)(p,h) variants. The notation (p,h)(p,h) follows the target definition in Section 3.2. For a forecast origin o, h is the forecast horizon, so the forecast is evaluated at the future endpoint o+ho+h. The parameter p is the comparison window ending at that endpoint: relative-change metrics use yo+h−py_o+h-p as the reference value and evaluate ro(p,h)=yo+h/yo+h−p−1r_o^(p,h)=y_o+h/y_o+h-p-1. Thus, p determines the economic movement being measured, while h determines how far ahead that movement is evaluated. Diagonal variants such as (1m,1m)(1m,1m) or (1y,1y)(1y,1y) measure the change from the forecast origin to the target endpoint, which is the natural setting for asset-like returns, cross-sectional ranking, and portfolio horizons. Non-diagonal variants such as (1y,1m)(1y,1m) measure a longer-window quantity at a shorter forecast horizon; for example, they evaluate the year-over-year value that will be observed one month after the forecast origin. In the compact IDs below, 1w1w, 1m1m, 1q1q, 1h1h, and 1y1y denote one week, one month, one quarter, one half-year, and one year, respectively. The variant IDs are used only to keep the metric-assignment tables compact; the corresponding (p,h)(p,h) combinations are listed explicitly in Table 5. Table 5: Mapping from compact variant IDs to (p,h)(p,h) combinations. (p,h)(p,h) Variant (p,h)(p,h) combinations D5 (1w,1w),(1m,1m),(1q,1q),(1h,1h),(1y,1y)(1w,1w),(1m,1m),(1q,1q),(1h,1h),(1y,1y) M10 (1m,1m),(1q,1m),(1q,1q),(1h,1m),(1h,1q),(1h,1h),(1y,1m),(1y,1q),(1y,1h),(1y,1y)(1m,1m),(1q,1m),(1q,1q),(1h,1m),(1h,1q),(1h,1h),(1y,1m),(1y,1q),(1y,1h),(1y,1y) M4 (1m,1m),(1q,1m),(1h,1m),(1y,1m)(1m,1m),(1q,1m),(1h,1m),(1y,1m) Q6 (1q,1q),(1h,1q),(1h,1h),(1y,1q),(1y,1h),(1y,1y)(1q,1q),(1h,1q),(1h,1h),(1y,1q),(1y,1h),(1y,1y) Q3 (1q,1q),(1h,1q),(1y,1q)(1q,1q),(1h,1q),(1y,1q) A1 (1y,1y)(1y,1y) N/A N/A Table 6: Metric assignments and variant IDs for inter-country Foreign Exchange evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant FX Pair Business-daily 34,503 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Table 7: Metric assignments and variant IDs for inter-country Commodity evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Future Front Business-daily 43,601 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Spot Business-daily 4,932 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Spot Monthly 975 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Inventory Stock Weekly 2,303 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Commodity ETF Single Business-daily 8,245 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Commodity ETF Broad Business-daily 3,298 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Commodity ETF Thematic Business-daily 3,298 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Table 8: Metric assignments and variant IDs for inter-country Crypto evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Crypto Pair Close (Top 100) Daily 120,963 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Table 9: Metric assignments and variant IDs for inter-country Market Index evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Country Index Business-daily 31,391 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Regional Index Business-daily 3,304 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Table 10: Metric assignments and variant IDs for country-level Fixed Income evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Treasury Yield Business-daily 18,161 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Credit Spread Business-daily 5,230 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Bond Total Return Index Business-daily 5,230 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Money Market Business-daily 12,543 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A TIPS Yield Business-daily 6,604 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Breakeven Business-daily 3,304 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Corporate Yield Business-daily 3,302 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Term Spread Business-daily 3,304 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Forward Inflation Business-daily 1,652 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Inflation Expectation Monthly 530 Point-wise HR, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Breakeven Monthly 150 Point-wise HR, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Corporate Yield Monthly 150 Point-wise HR, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Mortgage Rate Weekly 812 Point-wise HR, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Table 11: Metric assignments and variant IDs for country-level Market Indicators evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Implied Volatility Business-daily 13,361 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A US Equity Index Business-daily 14,868 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Table 12: Metric assignments and variant IDs for country-level Macro Indicators evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Inflation Monthly 675 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Labor Monthly 1,046 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A GDP Quarterly 50 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Housing Price Monthly 747 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Productivity Quarterly 72 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Money Monthly 375 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Industrial Production by Industry Monthly 2,100 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Labor by Industry Monthly 1,200 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Activity Monthly 744 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Capacity Utilization by Industry Monthly 825 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Inventory Monthly 518 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Spending Monthly 450 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Energy Activity Monthly 397 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Sentiment Monthly 200 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Housing Monthly 219 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Income Monthly 225 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Credit Monthly 148 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Regional Fed Monthly 152 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Saving Monthly 75 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Stress Monthly 75 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Freight Monthly 73 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Portfolio Flows Quarterly 168 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Current Account Quarterly 48 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A TIC Flows Quarterly 95 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Financial Flows Quarterly 24 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A External Monthly 444 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A External Quarterly 75 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Banking Monthly 150 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Fiscal Monthly 75 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs M10 Point-wise Δ , Δ2 ^2HR M4 Point-wise MASEoriMASE_ori N/A Fiscal Quarterly 24 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Fiscal Annual 8 Point-wise HR, Δ , Δ2 ^2HR, MASErelMASE_rel, MASEabsMASE_abs A1 Point-wise MASEoriMASE_ori N/A International Investment Position Quarterly 120 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A International Investment Position Annual 4 Point-wise HR, Δ , Δ2 ^2HR, MASErelMASE_rel, MASEabsMASE_abs A1 Point-wise MASEoriMASE_ori N/A FDI Flows Quarterly 24 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A FDI Flows Annual 4 Point-wise HR, Δ , Δ2 ^2HR, MASErelMASE_rel, MASEabsMASE_abs A1 Point-wise MASEoriMASE_ori N/A Table 13: Metric assignments and variant IDs for individual-series ETF evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant ETF OHLCV Close Business-daily 3,919,924 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Table 14: Metric assignments and variant IDs for individual-series Equity evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Equity OHLCV Close Business-daily 7,743,280 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs D5 Point-wise MASEoriMASE_ori N/A Cross-sectional IC D5 Portfolio return, vol, Sharpe, MDD D5 Table 15: Metric assignments and variant IDs for individual-series Fundamentals evaluation categories. Detail Category Frequency Eval. Points Task Metrics (p,h)(p,h) Variant Capital Expenditures Quarterly 60,843 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Cash and Cash Equivalents Quarterly 84,757 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A EPS Basic Quarterly 75,581 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Cross-sectional IC Q6 EPS Diluted Quarterly 74,206 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Cross-sectional IC Q6 Long-Term Debt Quarterly 43,229 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Net Income Quarterly 80,909 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Cross-sectional IC Q6 Operating Cash Flow Quarterly 83,997 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Operating Income Quarterly 67,678 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Cross-sectional IC Q6 Revenue Quarterly 60,974 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Cross-sectional IC Q6 Stockholders Equity Quarterly 93,125 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Total Assets Quarterly 91,002 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Total Liabilities Quarterly 77,746 Point-wise HR, MASErelMASE_rel, MASEabsMASE_abs Q6 Point-wise Δ , Δ2 ^2HR Q3 Point-wise MASEoriMASE_ori N/A Appendix C Additional Result Tables This appendix provides three complementary views of model performance. Table 16 reports the overall leaderboard used in the main analysis. It shows each model’s overall FinVerse rank, computed as the sum of the point-wise, IC, and portfolio placement ranks, together with the aspect-level ranks and the corresponding K values. The K values are geometric means of metric-level ranks within each evaluation aspect. Table 17 reports scope-wise performance. It aggregates metric-level ranks separately within the Country, Individual, and Inter-country scopes, showing whether a model performs consistently across different financial data scopes or is relatively stronger in a particular scope. Table 18 reports category-wise performance. It further breaks down rank aggregation by financial category, such as Foreign Exchange, Commodity, Crypto, Fixed Income, Macro Indicators, ETF, Equity, and Fundamentals. This table highlights which models are preferred for specific financial data categories and shows that strong overall performance does not necessarily imply uniformly strong performance across all categories. Table 16: Overall and aspect-wise model rankings. Overall is the sum of the point-wise, IC, and portfolio placement ranks. K columns report the geometric mean of metric ranks within each aspect. Rank Model Overall Rank Point IC Portfolio Rank K Rank K Rank K 1 Reverso Small 10 5 9.02 1 8.16 4 12.10 2 Chronos-2 Synth 11 8 11.02 2 8.17 1 9.18 3 TiRex 1.1 17 1 6.42 3 9.36 13 14.91 4 TimesFM 2.5 200M 29 6 9.45 4 10.24 19 16.44 5 Chronos-2 30 7 9.50 13 15.39 10 14.55 6 Chronos-Bolt Base 31 10 12.19 15 15.67 6 13.81 7 Chronos-Bolt Small 36 11 12.20 11 14.98 14 15.20 8 T0 Alpha 39 12 12.83 24 18.10 3 10.60 9 TiRex 2 Pretrain 39 2 7.02 5 11.14 32 19.53 10 Toto 2.0 2.5B 39 17 14.97 14 15.52 8 14.23 11 TiRex 2 ZS 40 3 7.94 6 11.19 31 19.47 12 VisionTS 44 9 11.47 7 12.50 28 18.62 13 PatchFM 45 4 8.46 12 15.21 29 18.66 14 Toto 2.0 4M 48 14 13.70 22 17.39 12 14.90 15 TimesFM 2.0 500M 49 38 28.43 9 13.25 2 10.33 16 Toto 2.0 1B 54 15 14.22 18 16.69 21 16.89 17 Kairos-50M 55 18 15.14 28 19.09 9 14.46 18 TempoPFN 55 13 13.31 16 16.06 26 17.99 19 Granite FlowState R1 60 20 16.23 25 18.14 15 15.56 20 TimesFM 1.0 200M 62 21 16.32 17 16.07 24 17.89 21 Super Linear 63 23 17.40 10 14.77 30 19.15 22 Kairos-10M 65 27 19.96 27 18.90 11 14.78 23 Toto 2.0 22M 65 19 15.17 26 18.18 20 16.60 24 FlowState 67 22 16.72 29 19.36 16 15.66 25 Toto 2.0 313M 68 16 14.59 30 19.57 22 17.22 26 Timer-S1 70 24 17.75 8 13.16 38 22.27 27 Toto Open Base 1.0 74 26 19.19 31 20.38 17 16.21 28 Moirai 1.1-R Large 76 37 27.03 34 21.09 5 12.21 29 Moirai 1.1-R Base 80 36 26.64 37 23.38 7 13.98 30 Moirai 1.1-R Small 82 32 24.88 32 20.55 18 16.33 31 PatchTST-FM R1 83 33 25.45 23 17.52 27 18.54 32 Sundial Base 128M 85 25 19.02 20 17.34 40 22.75 33 Kairos-23M 87 28 20.69 36 23.31 23 17.80 34 Granite PatchTST-FM R1 91 39 28.55 19 16.94 33 19.56 35 Moirai 2.0-R Small 97 30 23.08 33 20.77 34 20.56 36 TTM R3 97 34 26.32 21 17.37 42 23.64 37 TTM R1 99 31 24.00 43 26.57 25 17.93 38 CleanTS-65M 105 35 26.36 35 22.21 35 20.69 39 TTM R2 114 29 21.32 42 26.51 43 24.51 40 YingLong 50M 116 41 34.31 38 23.58 37 21.77 41 YingLong 300M 118 43 34.61 39 23.70 36 21.21 42 YingLong 110M 121 42 34.53 40 23.93 39 22.58 43 YingLong 6M 122 40 33.99 41 24.32 41 23.08 Table 17: Scope-wise rank aggregation. Lower values are better. Model Country Individual Inter-country Reverso Small 9.21 9.03 9.12 Chronos-2 Synth 12.25 9.27 6.80 TiRex 1.1 5.96 8.21 11.43 TimesFM 2.5 200M 9.75 9.48 10.66 Chronos-2 9.08 13.40 11.33 Chronos-Bolt Base 12.02 14.18 12.56 Chronos-Bolt Small 11.99 15.72 11.92 T0 Alpha 11.38 15.90 18.32 TiRex 2 Pretrain 6.47 11.09 11.60 Toto 2.0 2.5B 15.73 15.78 11.02 TiRex 2 ZS 7.39 10.92 13.22 VisionTS 11.08 10.76 18.33 PatchFM 7.52 17.39 11.86 Toto 2.0 4M 13.28 19.24 12.41 TimesFM 2.0 500M 30.24 14.62 21.80 Toto 2.0 1B 14.82 16.30 11.43 Kairos-50M 14.67 18.81 14.67 TempoPFN 14.59 17.61 7.76 Granite FlowState R1 14.95 25.95 15.11 TimesFM 1.0 200M 16.40 12.99 20.85 Super Linear 18.47 11.79 19.42 Kairos-10M 19.61 24.96 14.96 Toto 2.0 22M 15.41 18.67 12.45 FlowState 15.29 28.62 15.00 Toto 2.0 313M 14.96 17.85 12.22 Timer-S1 18.53 12.78 20.72 Toto Open Base 1.0 19.41 20.09 16.32 Moirai 1.1-R Large 26.93 20.80 23.61 Moirai 1.1-R Base 26.45 21.43 24.92 Moirai 1.1-R Small 24.97 19.98 24.08 PatchTST-FM R1 26.15 17.60 25.66 Sundial Base 128M 19.20 16.09 22.81 Kairos-23M 19.94 23.51 20.88 Granite PatchTST-FM R1 29.88 17.93 27.28 Moirai 2.0-R Small 22.06 24.82 24.66 TTM R3 28.33 16.70 25.25 TTM R1 23.88 20.86 25.44 CleanTS-65M 26.34 21.88 27.44 TTM R2 21.91 18.75 23.87 YingLong 50M 36.21 24.99 27.17 YingLong 300M 36.40 25.00 27.61 YingLong 110M 36.21 25.69 28.06 YingLong 6M 35.89 24.89 27.83 Table 18: Frequency-wise rank aggregation. Lower values are better. Model Business-daily Daily Weekly Monthly Quarterly Annual Reverso Small 7.33 11.85 11.86 9.98 8.90 16.53 Chronos-2 Synth 7.09 4.30 8.20 11.50 15.17 14.42 TiRex 1.1 10.24 14.85 8.36 4.60 10.11 12.45 TimesFM 2.5 200M 9.91 16.45 9.34 9.28 10.64 8.19 Chronos-2 12.33 12.36 10.20 8.66 10.45 18.61 Chronos-Bolt Base 12.59 12.50 7.93 12.94 11.42 13.36 Chronos-Bolt Small 11.57 14.08 9.53 12.70 12.92 13.28 T0 Alpha 18.12 21.50 20.42 9.15 17.35 22.63 TiRex 2 Pretrain 13.24 9.67 10.89 4.99 10.73 11.18 Toto 2.0 2.5B 9.33 12.90 7.63 16.37 19.58 33.25 TiRex 2 ZS 14.79 13.28 10.46 6.01 10.13 10.96 VisionTS 18.96 23.78 28.45 10.54 8.90 9.46 PatchFM 9.89 15.21 10.66 6.63 15.34 17.73 Toto 2.0 4M 12.59 12.96 18.29 13.19 16.57 22.87 TimesFM 2.0 500M 21.97 13.15 25.56 34.18 18.06 14.07 Toto 2.0 1B 10.92 12.81 6.58 14.59 18.99 35.90 Kairos-50M 12.03 20.66 15.61 17.92 13.02 15.20 TempoPFN 6.40 21.72 7.97 16.24 18.87 19.57 Granite FlowState R1 15.07 22.86 14.19 15.41 18.87 23.77 TimesFM 1.0 200M 20.47 19.90 26.10 15.53 14.53 11.31 Super Linear 20.61 21.80 24.40 18.26 13.00 13.54 Kairos-10M 13.77 19.51 24.40 22.18 20.17 26.28 Toto 2.0 22M 13.14 10.04 13.58 16.04 16.57 33.38 FlowState 14.63 24.36 14.30 15.41 21.93 20.51 Toto 2.0 313M 12.64 11.57 8.35 14.41 19.32 31.23 Timer-S1 23.74 21.87 11.05 17.16 15.04 13.89 Toto Open Base 1.0 17.92 13.24 16.23 18.81 21.39 21.30 Moirai 1.1-R Large 21.67 11.44 16.15 28.50 26.13 12.48 Moirai 1.1-R Base 23.26 17.71 17.39 27.20 25.13 28.30 Moirai 1.1-R Small 23.76 18.44 21.73 26.65 20.14 21.77 PatchTST-FM R1 27.78 21.49 24.39 24.66 22.07 25.02 Sundial Base 128M 25.88 18.68 16.51 17.87 17.28 12.67 Kairos-23M 21.41 23.28 19.42 21.09 18.82 15.77 Granite PatchTST-FM R1 29.63 21.82 32.35 29.65 21.63 23.38 Moirai 2.0-R Small 23.40 20.47 25.18 23.33 21.80 11.92 TTM R3 25.16 27.01 34.44 28.75 20.55 19.58 TTM R1 26.85 22.99 33.17 29.33 13.76 4.01 CleanTS-65M 29.28 22.78 35.95 24.96 24.33 24.97 TTM R2 23.61 26.08 36.03 28.45 11.17 6.05 YingLong 50M 29.83 22.11 34.05 36.50 29.89 27.14 YingLong 300M 30.70 18.62 34.37 36.64 30.10 27.14 YingLong 110M 30.49 21.04 34.42 36.55 30.43 27.14 YingLong 6M 29.31 22.38 34.41 36.49 29.92 27.14