Paper deep dive
A Predict-then-Correct Loop Based on Few-Shot Continuous Contextual Bandit for Demand Forecasting
Zhiwei Lei, Benedict Jun Ma, Ilya Jackson
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/21/2026, 3:06:04 AM
Summary
This paper proposes a Predict-then-Correct (PtC) framework for retail demand forecasting that addresses the challenge of sparse feedback in early demand cycles. The framework retains a first-stage machine learning forecast and applies a few-shot continuous contextual bandit (FSCCB) correction policy. Key innovations include similar-SKU augmentation to handle data scarcity and a top-p masked updating mechanism to prevent catastrophic forgetting. Experimental results on Walmart and beverage datasets demonstrate statistically significant reductions in MAPE, MAE, and RMSE compared to ML-only baselines and other policies like base-stock, PPO, and SAC.
Entities (12)
Relation Signals (9)
Predict-then-Correct Framework → evaluatedon → Walmart Dataset
confidence 95% · Across Walmart retail data ... PtC delivers statistically significant reductions
Predict-then-Correct Framework → evaluatedon → Beverage Dataset
confidence 95% · Across ... an exclusive beverage dataset, PtC delivers statistically significant reductions
Predict-then-Correct Framework → uses → Few-Shot Continuous Contextual Bandit
confidence 95% · proposing a predict-then-correct (PtC) framework that retains a first-stage machine learning (ML) forecast and applies a few-shot continuous contextual bandit correction policy
Predict-then-Correct Framework → outperforms → ML-only baseline
confidence 92% · improves average RMSE by 9.52% over the ML-only baseline in the ablation study
Few-Shot Continuous Contextual Bandit → employs → Top-p Masked Updating
confidence 90% · applies a few-shot continuous contextual bandit correction policy with ... top-p masked updating
Few-Shot Continuous Contextual Bandit → employs → Similar-SKU Augmentation
confidence 90% · applies a few-shot continuous contextual bandit correction policy with similar-SKUs augmentation
Predict-then-Correct Framework → outperforms → Proximal Policy Optimization
confidence 90% · yields lower inventory costs than ... proximal policy optimization
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Retail demand forecasting remains difficult when demand shifts faster than static forecasting models can be retrained, especially in early demand cycles where newly observed labels are sparse. To address this, this study aims to improve adaptive retail forecasting by proposing a predict-then-correct (PtC) framework that retains a first-stage machine learning (ML) forecast and applies a few-shot continuous contextual bandit correction policy with similar-SKUs augmentation and top-p masked updating. Across Walmart retail data and an exclusive beverage dataset, PtC delivers statistically significant reductions in MAPE, MAE, and RMSE across stable & high volume, stable & low volume, and erratic & intermittent demand patterns, improves average RMSE by 9.52% over the ML-only baseline in the ablation study, and yields lower inventory costs than base-stock, proximal policy optimization, and soft actor-critic policies under the tested lead-time settings. These findings show that online forecast correction can bridge offline demand learning and real-time retail decision-making by adapting to sparse feedback without fully retraining the base forecasting model.
Tags
Links
- Source: https://arxiv.org/abs/2607.16354v1
- Canonical: https://arxiv.org/abs/2607.16354v1
Trouble viewing inline? Open PDF directly →
Full Text
85,284 characters extracted from source content.
Expand or collapse full text
A Predict-then-Correct Loop Based on Few-Shot Continuous Contextual Bandit for Demand Forecasting Zhiwei Leia, Benedict Jun Maa and Ilya Jacksonb aThrust of Intelligent Transportation, Hong Kong University of Science and Technology (Guangzhou), China; bCenter for Transportation and Logistics, Massachusetts Institute of Technology, USA; Abstract Retail demand forecasting remains difficult when demand shifts faster than static forecasting models can be retrained, especially in early demand cycles where newly observed labels are sparse. To address this, this study aims to improve adaptive retail forecasting by proposing a predict-then-correct (PtC) framework that retains a first-stage machine learning (ML) forecast and applies a few-shot continuous contextual bandit correction policy with similar-SKUs augmentation and top-p masked updating. Across Walmart retail data and an exclusive beverage dataset, PtC delivers statistically significant reductions in MAPE, MAE, and RMSE across stable & high volume, stable & low volume, and erratic & intermittent demand patterns, improves average RMSE by 9.52% over the ML-only baseline in the ablation study, and yields lower inventory costs than base-stock, proximal policy optimization, and soft actor-critic policies under the tested lead-time settings. These findings show that online forecast correction can bridge offline demand learning and real-time retail decision-making by adapting to sparse feedback without fully retraining the base forecasting model. Keywords: demand forecasting; predict-then-correct; contextual bandit; few-shot learning. 1 Introduction Retail supply chains are high-dimensional, uncertain, and complex decision systems that tightly couple demand forecasting, inventory control, and multi-stage logistics operations (Cohen et al., 2022). This system-level complexity is driven primarily by three interrelated dimensions: stochastic demand, network structure, and dynamic decision-making. First, customer demand is non-stationary and volatile, often affected by exogenous market conditions, pricing mechanisms, and promotional activities (Fahimnia et al., 2025). Second, managing millions of stock-keeping units (SKUs) across multi-echelon distribution networks introduces intricate product-level relationships, such as substitution effects, complementarity, and cross-item resource sharing (Ma et al., 2016). Third, key operational decisions, including inventory allocation (Khouja and Zhou, 2019; Kim et al., 2024), dynamic pricing (Zhong et al., 2020), and robust procurement (Wagner, 2015; Li et al., 2024a), exhibit complex time-dependent characteristics (Huang et al., 2019). Consequently, inaccurate demand forecasts can propagate beyond the forecasting stage and distort downstream operational decisions, leading to severe stockouts or costly backlogs that degrade overall supply chain performance (Cohen et al., 2022). With the rapid development of artificial intelligence (AI) and enterprise big data infrastructures, modern supply chains are actively evolving into data-driven cyber-physical systems that achieve real-time perception and dynamic closed-loop optimization (Punia and Shankar, 2022; Mandania and Oliveira, 2023). In this context, demand forecasting constitutes a fundamental analytical function for such intelligent supply chain systems. Industry evidence suggests that AI-driven forecasting frameworks can reduce prediction errors by 20% to 50%, lower holding costs by 5% to 10%, and reduce lost sales by up to 65% (McKinsey, 2022). Methodologically, this development has accelerated a shift in forecasting research from classical parametric time-series models toward supervised machine learning (ML) architectures, such as long short-term memory (LSTM) networks (Hochreiter and Schmidhuber, 1997), gradient boosting machines (GBM) (Friedman, 2001), XGBoost (Chen and Guestrin, 2016), and deep learning (DL) architectures such as transformer-based variants (Zhang et al., 2024; Li et al., 2024b). Due to the abilities of nonlinear mapping, these data-driven approaches have demonstrated decisive algorithmic superiority over traditional statistical benchmarks in large-scale forecasting (Makridakis et al., 2022; Lei et al., 2024). Despite these advances, offline-trained ML models remain limited when deployed in rapidly changing and non-stationary market environments. During a new demand cycle or a product launch period, market dynamics often suffer from sharp structural breaks compared to historical regimes, inducing acute data sparsity and prominent “cold-start” challenges (Petropoulos et al., 2022). Because traditional supervised pipelines rely on rigid, batch-oriented training and require massive data pools for expensive offline retraining, they cannot seamlessly ingest real-time demand feedback. In practice, the correction of forecast bias therefore remains largely reactive (Wang et al., 2024). Once a base model is trained and deployed, it operates as an open-loop predictor that lacks the capacity for real-time calibration against short-term operational fluctuations and concurrent temporal shifts. This reveals a critical methodological void: the absence of a systematic, sample-efficient mechanism engineered to explicitly learn to correct baseline prediction errors on the fly. To bridge this gap, this study proposes a novel predict-then-correct (PtC) loop framework designed for adaptive demand calibration. Instead of treating forecast-error correction as a passive label-fitting task, we reformulate real-time forecast adjustment as a contextual bandit (CB) problem. In the proposed framework, an offline ML model first generates baseline predictions, after which an online few-shot continuous contextual bandit (FSCCB) policy dynamically fine-tunes these predictions by selecting bounded, continuous corrective actions from streaming market contexts based on immediate reward feedback. The core advantage of this architecture lies in its decoupled, closed-loop design: it preserves the structural, historical knowledge captured by the offline base model, while enabling the system with real-time, step-by-step responsiveness as market observations gradually unfold. To resolve data scarcity and the stability-plasticity dilemma inherent in early-cycle launches, the framework further integrates a feature-similar sequence data augmentation strategy and a sparse parameter update mechanism (Mazumder et al., 2021) that updates only highly sensitive policy parameters while freezing the primary structural network. Ultimately, this framework shifts the operational forecasting paradigm from static prediction to autonomous, self-calibrating intelligent control. The main contributions of this paper are summarized as follows: (1) We propose a two-stage PtC framework that couples AI/ML forecasting with a CB correction model. This transforms static, open-loop demand forecasting into an adaptive, closed-loop system, significantly enhancing prediction responsiveness to market fluctuations. (2) To tackle real-time data scarcity, we develop a few-shot continuous updating strategy. A base model is pretrained on historical data for initialization, and then dynamically fine-tuned via streaming data augmentation and a selective parameter update mechanism, enabling rapid adaptation to new demand patterns while preventing overfitting and catastrophic forgetting. The remainder of the paper is organized as follows. Related works are reviewed in the next section. Section 3 describes the details of the two-stage framework. The experimental results are discussed in Section 4. Section 5 discusses the application of PtC in the inventory management system. Conclusions are presented in Section 6. 2 Literature review 2.1 Traditional statistical methods for demand forecasting Classical demand forecasting research first developed around extrapolative time-series models. Naive, moving-average, exponential-smoothing, and ARIMA-family models remain widely used because they are transparent, computationally efficient, and effective when demand contains stable level, trend, or seasonal components (Taylor, 2003; Nikolopoulos et al., 2011; Hyndman and Athanasopoulos, 2018). Exponential-smoothing models place greater weight on recent observations and therefore provide a simple way to adapt to gradual level or trend changes, while ARIMA-type models use autoregressive and moving-average structures to capture serial dependence after differencing. These models are especially attractive in operational environments because they are easy to implement and require limited feature engineering. Retail applications extended these ideas to volatile and skewed demand settings: early work on retailer demand forecasting identified promotions, competitor actions, weather, and holidays as important drivers (Geurts and Kelly, 1986), while later studies developed interval forecasts, hybrid SARIMAX quantitative methods, and intermittent-demand aggregation strategies for demand series with irregular arrivals or heavy-tailed errors (Taylor, 2007; Arunraj and Ahrens, 2015; Nikolopoulos et al., 2011). The strength of this stream is its clarity: the source of each forecast is usually traceable to level, seasonality, trend, or a small set of lagged terms. Its limitation is equally clear. Once the functional form has been specified, the model has limited flexibility to represent nonlinear promotion effects, changing consumer response, cross-SKU substitution, or sudden shifts in demand regimes. Thus, classical statistical models provide reliable baselines, but they are less suited to settings where the main forecasting challenge is rapid adaptation rather than stable extrapolation. A second stream uses causal, multivariate, and econometric models to incorporate explanatory factors into demand forecasts. In retail systems, demand is rarely driven by time alone. All prices, discounts, display activities, holidays, competitive information, search intensity, weather, social-media signals, and product attributes can alter realized sales. Therefore, operations and forecasting studies have examined how external information can be integrated into forecast pipelines and how judgmental or promotional adjustments affect forecast quality (Fildes et al., 2008; Trapero et al., 2013; Steinker et al., 2017; Boone et al., 2018; Cui et al., 2018; Fildes et al., 2022). In fast-moving consumer goods and retail settings at the SKU-level, variable selection and structured econometric models have been used to address high-dimensional promotional variables, competitive effects, and time-varying marketing impacts (Ma et al., 2016; Huang et al., 2014, 2019; Ye et al., 2024). Hierarchical Bayesian and finite-mixture extensions further examine whether store-level heterogeneity improves forecast and elasticity estimation (Andrews et al., 2008). These studies make an important contribution by linking prediction to interpretable business mechanisms: instead of treating demand as an autonomous time series, they show how observed operational levers can improve forecast accuracy and managerial understanding. However, this gain in explanatory richness creates a practical tradeoff. Causal and multivariate models require well-measured covariates, stable relationships between covariates and demand, and enough observations to estimate the effects reliably. When many promotional, price, calendar, and competitive variables are included, the model can face sparse event coverage and a heavy variable-selection burden (Trapero et al., 2015). These difficulties are particularly acute for new products, short sales windows, and low-volume SKUs, where the data needed to estimate a rich explanatory model may not yet exist. Large-scale empirical evidence also warns that model complexity does not automatically translate into better forecasts. The M-competitions showed that statistically sophisticated methods may fail to outperform simpler alternatives on broad forecasting benchmarks (Makridakis et al., 1982, 1993; Makridakis and Hibon, 2000), and related evidence argues that excessive model complexity can reduce accuracy, increase error opportunities, and weaken managerial usability (Green and Armstrong, 2015). This finding is important for retail supply chains because forecasting is not an isolated statistical exercise. Forecasts are inputs to replenishment, allocation, pricing, and inventory-control decisions, so a method that is marginally more accurate but difficult to maintain may still be unattractive in practice. Studies in fast-fashion inventory and online-retail analytics illustrate that demand forecasts create value only when they are timely and operationally usable (Caro and Gallien, 2010; Ferreira et al., 2016). A forecast produced after a long retraining cycle may be too late for ordering decisions; a model that requires extensive manual feature redesign may be too costly for thousands of SKUs; and a method that performs well on average may still fail when early-cycle observations are sparse. 2.2 Machine learning and deep learning models ML and DL methods have expanded the forecasting toolkit by allowing nonlinear mappings, high-dimensional covariates, and large-scale pattern learning. Review studies report a broad adoption of ML tools in supply-chain demand forecasting, supplier selection, order allocation, and related operational prediction tasks (Aamer et al., 2021; Ingle et al., 2021; Islam et al., 2021; Feizabadi, 2022; Malviya and Bhandari, 2024). In retail demand prediction, richer information sources and supervised-learning methods have enabled models to use weather, promotion, search, social-media, product-attributes and price information on a scale that is difficult for traditional statistical models (Steinker et al., 2017; Boone et al., 2018; Cui et al., 2018; Lei et al., 2023; Ferreira et al., 2016; Alley et al., 2023). The 2020 M5 Accuracy competition further illustrates this shift: leading methods were ML-based and achieved strong accuracy improvements over traditional statistical benchmarks on large-scale hierarchical Walmart sales data (Makridakis et al., 2022). This line of work shows that data-intensive models can represent complex retail patterns, pool signals across products, and exploit high-dimensional covariates. It also changes the role of forecasting in supply chains. Forecasting is no longer only an extrapolation task based on a single SKU history; it becomes a pattern-learning problem that uses cross-product, temporal, promotional, and contextual signals to infer future demand. Within this broad family, different model classes address different sources of complexity. Tree-based ensemble models, such as gradient boosting frameworks, are effective when forecasting is based on tabular covariates, nonlinear interactions, and heterogeneous feature effects. Neural networks and deep sequence models capture nonlinear relationships and temporal dependencies. Early evidence of ANN, CNN-based time-series models, LSTM architectures, attention mechanisms, and Transformer variants have been used to improve retail or time-series forecasting (Adya and Collopy, 1998; Alon et al., 2001; Aburto and Weber, 2007; Zhang et al., 2019; Graves and Schmidhuber, 2005; Mamdouh et al., 2024; Gao et al., 2023; Zhang et al., 2024; Li et al., 2024b). These models are attractive because they can learn representations rather than requiring researchers to specify all interactions manually. In the retail setting, this is especially useful when demand is jointly affected by price, promotion, calendar, assortment, and historical sales signals. At the same time, deep models can be data-hungry, sensitive to distribution shifts, and difficult to interpret. Their empirical success often depends on the availability of large training histories and on whether future conditions resemble the historical regimes used for training. When demand changes after deployment, simply relying on the offline model may be insufficient, even if the model was strong on the original test set. Other ML streams complement deep sequence forecasting by addressing uncertainty, dependency, and robustness. Gaussian process and neural network applications in agricultural, commodity, real-estate, and financial price series demonstrate the value of flexible nonlinear learning and uncertainty-aware prediction in related time-series environments (Xu and Zhang, 2021, 2023; Jin and Xu, 2025b). Graph and causal structure methods address interconnected markets and contemporaneous dependencies, which are relevant to retail systems where products may substitute, complement, or share demand drivers (Jin and Xu, 2025a). Ensemble and composite forecasts improve robustness by combining multiple base learners and reducing dependence on a single model specification (Guo et al., 2025). These approaches broaden the modeling landscape, but they also highlight a recurring limitation. However, the literature has developed many powerful first-stage predictors, but the second-stage question of real-time forecast correction remains less developed. The need for adaptation has motivated online learning, reinforcement learning, and continual-learning approaches in operations and forecasting. Online-learning studies emphasize updating models as new data arrive, while reinforcement-learning studies show how operational policies can learn from repeated interaction with uncertain environments (Petropoulos et al., 2022; Wang et al., 2024; Boute et al., 2022; Gijsbrechts et al., 2022). These studies are important because they move beyond purely static prediction and recognize that decisions and observations unfold sequentially. However, a direct reinforcement-learning formulation may be excessive for demand correction when the immediate goal is not to learn an entire inventory policy but to adjust a baseline forecast before downstream decisions are made. Similarly, online retraining can be costly and unstable when only a few new observations are available. Continual-learning and few-shot studies further show that unrestricted parameter updates can cause overfitting or catastrophic forgetting under limited new data (Kirkpatrick et al., 2017; Tian et al., 2024; Bethune et al., 2025; Mazumder et al., 2021). This creates a stability-plasticity dilemma: a model must adapt enough to correct new bias, but not so much that it destroys the useful structure learned from previous demand cycles. The closest unresolved issue is therefore not whether advanced forecasting models can be accurate offline, but how their forecasts should be adapted online after new demand feedback is observed. Most demand forecasting pipelines still generate static predictions, select among static models, or require costly retraining when demand changes. They rarely treat each forecast as a decision that can be immediately corrected using contextual feedback. This distinction is central to our study. The PtC framework does not attempt to replace statistical, machine-learning, or deep-learning forecasters. Instead, it uses them as the first stage and adds a CB-based correction policy as the second stage. The CB formulation is appropriate because each correction action receives immediate feedback through the change in forecast error, which makes the learning signal local, interpretable, and sample efficient. The few-shot continuous updating design further addresses the limited-data setting by updating only selected parameters through the Top-p mechanism, thus preserving prior correction knowledge while allowing local adaptation. 3 Solution methods 3.1 AI/ML prediction and data augmentation Figure 1 illustrates the overall framework of the proposed PtC method. In the first stage, using average demand, coefficient of variation (CoV), and intermittency as clustering features, SKUs will be divided into four categories: stable & high volume, stable & low volume, erratic & intermittent, and lumpy (Ma et al., 2025). Each category is then associated with an ML algorithm for prediction. For the target SKU, the first step is to identify its category and then generate a baseline forecast using the corresponding ML model. The resulting baseline forecast is used as input to the second stage. Augmentation data selection within each cluster is provided in Appendix Appendix C: Augmentation Data Selection. Figure 1: Total framework of the proposed method. In the second stage, we adopt the CB as the correction model to adjust the ML-predicted demand by real-time observations within a sliding window. To support timely prediction and avoid delaying operational decisions, only five or seven days of demand information can be collected; otherwise, the correction loses its practical relevance. In this setting, real-time demand information from a single SKU is insufficient to train an effective correction policy and may cause severe overfitting or catastrophic forgetting. Therefore, it is necessary to design a data augmentation strategy to expand the training set and alleviate these risks. For the target SKU, there are often numerous SKUs within the same cluster that share similar characteristics. Thus, we first rank candidate SKUs by their Euclidean distance to the target SKU and then select the k nearest SKUs to augment the training set. Meanwhile, an upper distance threshold disd_is is set to avoid selecting SKUs that are too dissimilar from the target SKU. If fewer than k SKUs satisfy this threshold, we apply same-trend perturbations to both historical demand and ML-predicted demand until a pre-specified minimum training-set size is reached. Clustering, scaling, neighbor selection, and augmentation parameters are estimated using training-period information only to prevent test-period leakage. 3.2 FSCCB 3.2.1 State, action, and reward design With full observation enabled, the observed state, or context, of the PtC system at time step t is denoted as: st=grd,gxd,ged,ghrd,ghxd,gsd,gdt,grt,gwt,ght,gst.s_t= \g^d_r,g^d_x,g^d_e,g^d_hr,g^d_hx,g^d_s,g^t_d,g^t_r,g^t_w,g^t_h,g^t_s \. (1) The description of each component is shown in Table 1. Specifically, grdg^d_r represents the historical real demand, gxdg^d_x represents the historical ML-predicted demand, and gedg^d_e represents the historical prediction error. In addition, ghrdg^d_hr and ghxdg^d_hx denote the mean historical real demand and the mean historical ML-predicted demand, respectively, while gsdg^d_s measures the historical slope of the real demand. Time-related features include the current time step, the remaining time before the end of the cycle, and binary indicators for weekend, holiday and shopping festival effects. Table 1: State representation components. Type Description Notation Historical real demand grdg_r^d Historical ML-predicted demand gxdg_x^d Historical data Historical ML error gedg_e^d Mean historical real demand ghrdg_hr^d Mean historical ML-predicted demand ghxdg_hx^d Historical slope of real demand gsdg_s^d Current time step gdtg_d^t Remaining time before the end of the cycle grtg_r^t Time characteristics Is weekend or not gwtg_w^t Is holiday or not ghtg_h^t Is shopping festival or not gstg_s^t In the CB framework, sts_t is used as the contextual information available before making the correction decision. The policy network uses this context to determine how the original ML prediction should be adjusted. In this study, the action is defined as the continuous adjustment factor applied to the ML-predicted demand. Given the observed context sts_t, the policy network outputs the mean and standard deviation of a Gaussian correction distribution, denoted by μθ(st) _θ(s_t) and σθ(st) _θ(s_t), respectively. During training and online adaptation, the correction action is sampled as a~t∼(μθ(st),σθ2(st)),at=clip(a~t,−1,2). a_t ( _θ(s_t), _θ^2(s_t) ), a_t=clip ( a_t,-1,2 ). (2) Here, ata_t denotes the sampled relative correction ratio. A negative value indicates that the original ML prediction should be reduced, while a positive value indicates that the original ML prediction should be increased. The action range is limited to [−1,2][-1,2] to prevent the correction model from generating unreasonable adjustment values. The stochastic sampling in Eq. (2) provides local exploration over possible upward and downward forecast adjustments. Given the original ML-predicted demand y^tML y^ML_t, the corrected demand is calculated as: y^tCB=y^tML(1+at∗), y^CB_t= y^ML_t(1+a_t^*), (3) where y^tCB y^CB_t denotes the demand value corrected by the CB model. Therefore, the CB policy does not directly predict the final demand. Instead, it learns how to adjust the baseline ML prediction according to the observed context. The reward is defined as the relative degree of improvement after correction. At each time step t, the prediction error of the original ML model is calculated as: eML,t=|y^tML−yt|,e_ML,t= | y^ML_t-y_t |, (4) where yty_t denotes the actual demand. Similarly, the prediction error after the CB correction is calculated as: eCB,t=|y^tCB−yt|.e_CB,t= | y^CB_t-y_t |. (5) To evaluate whether the correction action improves the original ML prediction, the reward function is designed as: rt=τ(eML,t−eCB,t),r_t=τ (e_ML,t-e_CB,t ), (6) where τ is a scaling factor used to normalize the reward into a suitable numerical range. According to Eq. (6), a positive reward indicates that the CB correction reduces the prediction error compared with the original ML prediction. A negative reward means that the correction worsens the prediction result. Therefore, maximizing the reward is equivalent to learning a correction policy that improves the forecasting accuracy of the baseline ML model. 3.2.2 CB for forecast correction In this study, the forecast correction problem is formulated as a continuous contextual bandit problem. The proposed PtC framework first uses the baseline ML model to generate an initial demand prediction. Then, the CB model samples a continuous correction action according to the stochastic policy πθ(⋅∣st) _θ(· s_t), the main steps are provided in Figure 2. The correction process can be represented as: st→at∗→rt,s_t→ a_t^*→ r_t, where sts_t denotes the observed context, at∗a_t^* denotes the correction action, and rtr_t denotes the reward measuring the improvement after correction. At each time step t, the correction model first observes the context sts_t, which is constructed from recent demand, baseline ML forecasts, historical forecast errors, and time-related features. The policy then samples a continuous correction action ata_t from πθ(⋅∣st) _θ(· s_t) and applies it to the baseline forecast. This stochastic action-selection mechanism enables exploration around the current policy mean and allows the model to test whether increasing, decreasing, or retaining the baseline forecast is more beneficial under the current context. Figure 2: Main steps of the second stage in PtC. The bandit feedback structure is essential in this formulation. Before the actual demand yty_t is observed, the model only has access to sts_t and must commit to one correction action ata_t. After yty_t is realized, the reward rtr_t is computed by comparing the baseline forecast error and the corrected forecast error. The model uses the reward of the selected action to update the policy, while no supervised target action is imposed. The newly observed tuple: (st,at,rt,logπθ(at∣st))(s_t,a_t,r_t, _θ(a_t s_t)) is then added to the sliding-window feedback set and used for future policy updates. Therefore, PtC learns from bandit feedback by increasing the probability of correction actions that receive positive rewards and decreasing the probability of actions that worsen the baseline forecast. 3.2.3 Learning Objective of the CB Correction Model Let the chronological feedback set be: =(st,at∗,rt)t=1T,D= \(s_t,a_t^*,r_t) \_t=1^T, (7) where each tuple contains the context, the correction action, and the realized reward. The goal is to learn a policy πθ _θ that maps the observed context to a continuous correction ratio: at∗=πθ(st).a_t^*= _θ(s_t). (8) Given the first-stage forecast y^tML y^ML_t, the corrected forecast induced by this policy is: y^tCB=y^tML(1+πθ(st)). y^CB_t= y^ML_t (1+ _θ(s_t) ). (9) Thus, the policy network does not replace the first-stage forecasting model. It only learns a context-dependent adjustment factor for the existing forecast. At the population level, the CB correction policy is trained to maximize the expected reward obtained after correction: πθ∗=argmaxπθst∼[rt(st,πθ(st))]. _θ^*= _ _θE_s_t [r_t(s_t, _θ(s_t)) ]. (10) This objective is appropriate for forecast correction because each action is evaluated by its immediate effect on the current prediction error. Unlike a full reinforcement-learning formulation, no long-horizon value function is required; the correction model only needs to learn whether the current context calls for increasing, decreasing or retaining the baseline forecast. In empirical training, the expected objective is approximated by the average reward over the feedback set: θ∗=argmaxθ1T∑t=1Trt.θ^*= _θ 1T _t=1^Tr_t. (11) Substituting the reward definition from Eq. (6) into Eq. (11) gives: θ∗=argmaxθ1T∑t=1Tτ(|y^tML−yt|−|y^tML(1+πθ(st))−yt|).θ^*= _θ 1T _t=1^Tτ ( | y^ML_t-y_t |- | y^ML_t (1+ _θ(s_t) )-y_t | ). (12) The first error term in Eq. (12) is determined by the baseline ML forecast and does not depend on θ. Therefore, maximizing the reward is equivalent to minimizing the corrected prediction error: ℒ(θ)=1T∑t=1T|y^tML(1+πθ(st))−yt|.L(θ)= 1T _t=1^T | y^ML_t (1+ _θ(s_t) )-y_t |. (13) This equivalence connects the reward-based contextual bandit formulation with the standard forecasting objective: a high reward means that the correction policy reduces the absolute error of the first-stage forecast. Proposition 1. Let LPtC∗L_PtC^* and LML_ML denote the optimal corrected and ML-only expected losses induced by any nonnegative forecasting loss ℓ . If 0∈0 and π0(s)=0∈Π _0(s)=0∈ , then LPtC∗≤LML_PtC^*≤ L_ML. Proposition 1 establishes a hypothesis-class containment result at the population optimum. Since the zero-correction policy π0(s)=0 _0(s)=0 is included in the PtC policy class, PtC contains the ML-only forecast as a special case, and its optimal population loss is no larger than that of the ML-only forecast. When the correction action is zero for every context, the corrected forecast exactly reduces to the original ML forecast. This result should be interpreted as an expressiveness and containment property of the PtC hypothesis class, rather than as a finite-sample performance guarantee: it does not imply that the estimated PtC model will necessarily outperform the ML-only baseline with limited samples or at every online decision step. A proof of Proposition 1 is provided in Appendix Appendix D: Proof of Proposition 1. In implementation, historical demand cycles are replayed in chronological order to initialize the policy, and the same feedback structure is used during online deployment. This design preserves the decision timing of the CB formulation: the policy observes sts_t, selects at∗a_t^*, receives rtr_t after demand is realized and then updates itself for future correction decisions. 3.2.4 Top-p update In a new demand cycle, customer demand is gradually observed over time. The CB correction model must update its parameters to incorporate new information from the demand environment. However, in such a data-scarce scenario, frequent parameter updates may lead to overfitting or catastrophic forgetting of previously learned knowledge (Kirkpatrick et al., 2017; Tian et al., 2024; Bethune et al., 2025). To mitigate this risk, we adopt a parameter update rule based on the proportion of parameters. We adopt a top-p masked update rule to control the degree of online adaptation. Specifically, we select the top-p proportion of parameters with the smallest absolute values and update only these parameters, while the remaining parameters are frozen. Small-magnitude parameter selection is a heuristic intended to preserve highly expressed pretrained parameters; its effectiveness must therefore be established empirically. Intuitively, parameters with larger absolute values are treated as more strongly expressed components of the pretrained correction policy, whereas smaller-magnitude parameters provide a limited adaptation subspace for learning new-cycle feedback. To implement this rule, we first sort the parameters in each layer by their absolute values. The threshold δkδ^k is then set to the largest magnitude within the smallest p%p\% parameters of layer k. Parameters whose absolute values are no larger than δkδ^k are selected for online updating, while the remaining parameters are kept fixed. After the threshold is determined, we adopt a binary mask matrix in equation (14) to record which parameters need to be updated as follows: M(θjk)=1,|θjk|≤δk0,otherwiseM( _j^k)= cases1,&| _j^k|≤δ^k\\ 0,&otherwise cases (14) where θjk _j^k is the j-th parameter in layer k. If M(θjk)=1M( _j^k)=1, the parameter θjk _j^k is selected to be updated. The shift of the update rule can be found in equation (15). θjk←θjk−ηgjk→θjk←θjk−ηM(θjk)gjk _j^k← _j^k-η\,g_j^k → _j^k← _j^k-η\,M( _j^k)\,g_j^k (15) where η represents the learning rate. In our research, we set the learning rate to a very small value to enable fine-tuning updates, thereby avoiding disruption of the knowledge learned before (Zhang et al., 2023; Tao et al., 2020). Proposition 2. Let Θp=θ0+Δ:supp(Δ)⊆Sp _p=\θ^0+ :supp( ) S_p\ and Sp1⊆Sp2S_p_1 S_p_2 for p1≤p2p_1≤ p_2. Then Gad(p)=infθ∈Θpℒnew(θ)−infθ∈Θ1ℒnew(θ)G_ad(p)= _θ∈ _pL_new(θ)- _θ∈ _1L_new(θ) is non-increasing in p. If |gju|≤G|g_j^u|≤ G, then ‖θpU−θ0‖2≤ηUG⌈pd⌉\| _p^U-θ^0\|_2≤η UG pd , so the forgetting-risk bound is non-decreasing in p. Proposition 2 characterizes the adaptation–stability tradeoff induced by the top-p update rule. The first part states that increasing p enlarges the set of trainable parameters. Since Θp1⊆Θp2 _p_1 _p_2 when p1≤p2p_1≤ p_2, a larger update set cannot increase the best achievable loss on the new demand cycle. Therefore, the adaptation gap Gad(p)G_ad(p) is non-increasing in p. This explains why updating too few parameters may limit the ability of the correction policy to adapt to a new demand pattern. proof of Proposition 2 is provided in Appendix Appendix E: Proof of Proposition 2. Proposition 3. Let ϵ(θ)=,at∼πθ[ρϵ(y^tML(1+at)−yt)]J_ε(θ)=E_D,a_t _θ[ _ε( y_t^ML(1+a_t)-y_t)] be lower bounded and L-smooth. Under θu+1=θu−ηmu⊙∇ϵ(θu)θ^u+1=θ^u-η\,m_u _ε(θ^u) with 0<η≤1/L0<η≤ 1/L, the PtC correction policy satisfies ϵ(θu+1)≤ϵ(θu)J_ε(θ^u+1) _ε(θ^u) for u=0,1,…u=0,1,…. Proposition 3 provides a stability guarantee for the masked update when the correction objective is represented by a smooth surrogate loss. The absolute forecasting error is not differentiable at zero, so ρϵ(⋅) _ε(·) is introduced as a smooth approximation of the correction error. If this surrogate objective is lower bounded and L-smooth, then the standard descent lemma implies that a masked gradient step with step size 0<η≤1/L0<η≤ 1/L does not increase the objective value. A proof of Proposition 3 is provided in Appendix Appendix F: Proof of Proposition 3. Pseudo-code of top-p masked update is described in Appendix Appendix G: Top-p Masked Bandit Update. 4 Experiment 4.1 Experimental design 4.1.1 Datasets and implementation All algorithms were implemented in Python 3.11 (64-bit). The experiments were conducted on a workstation equipped with an Intel Core i7-14700KF CPU @ 5.6 GHz, 32 GB DDR5 RAM, and an NVIDIA RTX 4070 Ti GPU. We evaluate the proposed PtC framework using the Walmart sales data released through the M5 Accuracy competition in 2020 (Makridakis et al., 2022). Specifically, we use 6,875 SKUs from Walmart stores “CA_1” to “CA_3”, covering more than 1,500 daily observations. The beverage dataset is an external industrial benchmark collected from a major North American beverage manufacturer. It contains 156 weeks of weekly shipment records 49 bottling plants and over 12,000 distribution nodes. The lumpy category is defined in Section 3 for completeness of SKU segmentation, but it is omitted from the later experiments because lumpy demand has long zero-demand periods and irregular spikes, making statistical and AI/ML models largely ineffective; therefore, we recommend a qualitative CPFR-based approach instead of reporting model-comparison results. Details of dataset splitting can be found in Appendix Appendix A: Dataset Splitting, and detailed descriptions of the parameter values of the set are provided in Appendix Appendix B: Principal Hyperparameter Settings. 4.1.2 Evaluation metrics We evaluate the precision of the forecast by mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean squared error (RMSE). For a test sequence with actual demand yty_t and forecast y^t y_t, the metrics are defined as: MAPE=100T∑t=1T|yt−y^tmax(|yt|,ϵ)|,MAPE= 100T _t=1^T | y_t- y_t (|y_t|,ε) |, (16) MAE=1T∑t=1T|yt−y^t|,RMSE=1T∑t=1T(yt−y^t)2.MAE= 1T _t=1^T|y_t- y_t|, RMSE= 1T _t=1^T(y_t- y_t)^2. (17) where ϵε is a small positive constant used to avoid division by zero. MAPE provides a scale-free percentage error, MAE reflects the average absolute deviation, and RMSE emphasizes large forecast errors. For the demand-category comparison, we report the relative error reduction: Reduction(M,q)=Eqref−EqMEqref×100%,Reduction(M,q)= E_q^ref-E_q^ME_q^ref× 100\%, (18) where EqME_q^M denotes the error of method M under metric q, and EqrefE_q^ref denotes the corresponding error of the ML-only reference forecast. A positive value indicates an improvement over the reference forecast, whereas a negative value indicates a deterioration in forecasting performance. 4.1.3 Benchmark methods ETS (Gardner Jr, 1985) is an exponential smoothing method that recursively updates level, trend, and seasonal components. A representative additive ETS forecast is given by y^∗t+hETS=ℓt+hbt+s∗t+h−m, y*t+h^ETS= _t+hb_t+s*t+h-m, (19) where ℓt _t, btb_t, and sts_t denote the level, and seasonal states, respectively, and m is the seasonal period. ARIMA is a classical linear time-series model that captures autocorrelation through autoregressive and moving-average terms after differencing. The model can be written as ϕ(B)(1−B)dyt=c+θ(B)εt,φ(B)(1-B)^dy_t=c+θ(B) _t, (20) where B is the backshift operator, d is the differencing order, and εt _t is the innovation term. LightGBM (Ke et al., 2017) is a gradient boosting decision tree model that builds an additive ensemble of regression trees. Its prediction can be expressed as y^∗t=∑∗k=1Kfk(xt),fk∈ℱ, y*t=Σ*k=1^Kf_k(x_t), f_k , (21) where xtx_t is the input feature vector and fk(⋅)f_k(·) denotes the k-th regression tree. XGBoost (Chen and Guestrin, 2016) is another boosted-tree model that learns additive regression trees by minimizing a regularized objective: ℒ(k)=∑t=1Tl(yt,y^t(k−1)+fk(xt))+Ω(fk),L^(k)= _t=1^Tl (y_t, y_t^(k-1)+f_k(x_t) )+ (f_k), (22) where l(⋅)l(·) is the loss function and Ω(fk) (f_k) penalizes tree complexity. N-BEATS is a deep neural forecasting model composed of stacked fully connected residual blocks. Each block generates a backcast for residual updating and a forecast, while the final prediction is obtained by summing the block-level forecasts. TiDE is a dense encoder–decoder forecasting model that maps historical demand and covariates into multi-step forecasts. TFT is an attention-based deep forecasting model that integrates static features, time-varying covariates, gating mechanisms, and temporal attention for multi-horizon forecasting. 4.2 Results 4.2.1 Numerical experiment on the Walmart dataset Table 2 reports the average relative error reduction on the Walmart retail dataset across three distinct demand categories. The results are calculated based on Eq. (18), where positive values indicate a performance improvement over the LightGBM ML-only reference forecast, and negative values denote an algorithmic decrease. Table 2: Comparison result of Average relative error reduction on the Walmart dataset. Method Stable & Low Volume Stable & High Volume Erratic & Intermittent MAPE MAE RMSE MAPE MAE RMSE MAPE MAE RMSE ETS 0.05 4.92 1.83 -9.66 -1.10 -2.08 -14.24 3.82 1.18 N-BEATS 7.31 5.23 1.42 -5.36 5.40 1.30 -26.85 5.19 -1.60 XGBoost -12.54 1.92 -0.81 -1.84 0.49 0.47 -8.78 -1.24 -13.75 ARIMA 1.03 4.68 1.90 -9.31 -4.17 -9.32 -14.77 3.67 1.09 TiDE 11.57 10.98 1.95 -5.82 5.17 0.86 -21.97 4.38 -1.29 TFT 6.67 10.78 0.86 -10.09 -0.43 -5.38 -25.38 4.81 -1.68 PtC 12.16*** 11.28*** 6.87*** 6.74*** 9.95*** 7.67*** 5.68*** 8.94*** 3.70*** Note: Positive values indicate improvement relative to the ML-only reference forecast, while negative values indicate performance decrease. * p<0.10p<0.10, ** p<0.05p<0.05, *** p<0.01p<0.01. Across all demand patterns, the proposed PtC framework consistently shows superior adaptability compared to both traditional statistical methods and DL architectures. Notably, PtC is the only method evaluated that achieves positive relative error reduction across all three metrics in every reported setting. Furthermore, all observed improvements are statistically significant at the p<0.01p<0.01 level. The empirical results expose the structural vulnerabilities of static baseline models when subjected to the highly stochastic nature of retail supply chains. Deep forecasting architectures, such as N-BEATS, TiDE, and TFT, exhibit moderate success under “Stable & Low Volume” conditions but degrade substantially when applied to “Erratic & Intermittent” demand. Specifically, N-BEATS shows a -26.85% decrease in MAPE for intermittent items. This underscores a limitation of pure data-driven models: their susceptibility to overfitting the noisy, sparse signals inherent in intermittent demand, which subsequently induces severe predictive bias. Conversely, traditional time-series methods like ETS and ARIMA struggle significantly with “Stable & High Volume” series. This is likely driven by their constrained parametric forms, which fail to capture the complex, non-linear dynamics characterizing high-throughput retail operations. In contrast, the performance of the PtC architecture achieves average error reductions between 5.68% and 8.94% even in the highly challenging erratic category, validating its underlying mathematical design. Rather than overriding the historical distributions captured by the base ML model, the FSCCB correction layer dynamically fine-tunes the forecast via online reward feedback. This decoupled architecture effectively bridges the methodological gap between offline historical generalization and online correction. 4.2.2 Numerical experiment on the beverage dataset To better validate the performance of the proposed correction mechanism, we conduct a supplementary numerical experiment on a beverage distribution dataset. Table 3 details the average relative error reduction under the previously established evaluation framework. Table 3: Comparison result of Average relative error reduction on the beverage dataset. Method Stable & Low Volume Stable & High Volume Erratic & Intermittent MAPE MAE RMSE MAPE MAE RMSE MAPE MAE RMSE LightGBM 6.18 5.05 6.12 – – – 3.91 -1.01 2.93 ETS – – – 7.82 2.89 -3.67 – – – N-BEATS 5.44 6.25 5.45 10.88 -5.57 -11.48 -10.39 4.01 2.20 XGBoost -5.94 4.71 -5.59 -6.24 -7.62 -8.46 0.79 0.64 2.79 ARIMA 6.80 -5.01 5.37 7.85 -18.07 -21.99 6.27 4.99 1.72 TiDE 6.91 5.43 5.58 11.19 -3.46 -6.14 1.24 -12.92 -0.04 TFT 5.49 6.76 -6.00 8.19 5.88 2.73 -10.33 7.06 2.27 PtC 13.64*** 7.85*** 7.09*** 14.64*** 14.85*** 13.40*** 8.92*** 12.37*** 4.48*** Note: Positive values indicate improvement relative to the ML-only reference forecast, while negative values indicate a performance decrease. * p<0.10p<0.10, ** p<0.05p<0.05, *** p<0.01p<0.01. In this data environment, traditional methods like ARIMA and tree-based models like XGBoost exhibit severe performance decrease in the “Stable & High Volume” category. Furthermore, DL models such as N-BEATS, TFT, yield highly inconsistent results across different metrics within identical demand categories. Such metric instability indicates that these base models are highly sensitive to shifting data distributions and lack the continuous learning capability required to mitigate real-time prediction errors under structural volatility. Conversely, PtC maintains substantial, balanced, and statistically significant (p<0.01p<0.01) improvements among all conditions. The error reductions peak in the “Stable & High Volume” category, achieving 14.64%, 14.85%, and 13.40% error reduction for MAPE, MAE and RMSE, respectively. By retaining the base ML model’s structural understanding of baseline demand while deploying an online CB to correct for micro-level volatility, PtC shifts the forecasting paradigm from a static, open-loop prediction task to an adaptive, closed-loop correction system. 4.3 Sensitivity analysis We further analyze the sensitivity of the top-p parameter update proportion and the depth of the correction network. Figure 3 compares the sensitivity of the update proportion p on the North American beverage company dataset and the Walmart retail dataset. The two curves show that updating too few parameters limits the performance of the correction model, while updating too many parameters increases the risk of overfitting or forgetting previously learned demand patterns. On the beverage-company dataset, the improvement rises rapidly from a small update proportion and reaches its peak p=0.16p=0.16, after which performance declines steadily. On the Walmart dataset, the improvement is already significant under small update proportions and reaches its highest value p=0.10p=0.10; further increasing p leads to a gradual decrease in improvement. This cross-dataset evidence suggests that FSCCB benefits from selective adaptation rather than full-network updating. (a) North American beverage company dataset (b) Walmart retail dataset Figure 3: Parameter sensitivity analysis of the update proportion p across two datasets. Figure 4 reports the sensitivity to network depth. A shallow correction network has limited representation capacity and may fail to capture nonlinear relationships between context features and correction actions. However, an overly deep correction network increases the number of trainable parameters and becomes less stable under few-shot updates. The observed pattern supports the use of a moderate-depth policy network for the CB correction model. Figure 4: Network depth sensitivity analysis. 4.4 Ablation study To isolate the contribution of the CB correction layer and the few-shot selective update rule, we conduct an ablation study using N-BEATS as the first-stage ML forecast. We compare three variants as follows: (1) ML-only: Forecasting relies on the N-BEATS prediction without second-stage correction; (2) ML+CB: A CB correction model is applied, but all correction-network parameters are updated during adaptation; (3) ML+FSCCB: FSCCB is applied, where only the selected top-p parameters are updated during adaptation. The reported Gap is the relative RMSE reduction compared with the ML-only baseline. A positive Gap indicates that the correction module improves the N-BEATS forecast. Table 4 reports the average ablation results of the proposed few-shot continuous CB method. Compared with the original N-BEATS model, the standard N-BEATS-CB only achieves a marginal RMSE reduction of 0.34%. Moreover, although its average RMSE is slightly lower than that of N-BEATS, its average MAE becomes higher. This indicates that simply introducing a CB-based correction network does not provide a sufficiently stable improvement under limited observations from a new demand cycle. Table 4: Average prediction performance comparison of N-BEATS, N-BEATS-CB, and PtC. N-BEATS N-BEATS-CB PtC RMSE MAE RMSE MAE Gap RMSE MAE Gap 11.4086 8.7993 11.3903 8.8363 0.34% 10.3372 7.8193 9.52% By contrast, PtC achieves a much larger average improvement. It reduces the average RMSE from 11.4086 to 10.3372 and the average MAE from 8.7993 to 7.8193, corresponding to an average RMSE Gap of 9.52%. This result shows that the proposed FSCCB method can more effectively adapt the forecasting model to a new demand cycle. The improvement mainly comes from the few-shot selective update mechanism. Instead of updating all parameters of the correction network, FSCCB updates only a small subset of low-magnitude parameters while freezing the remaining parameters. This strategy helps preserve the general correction knowledge learned from historical demand cycles and prevents the model from overfitting short-term noise in the new cycle. Therefore, the ablation results confirm that FSCCB provides a more stable and effective correction mechanism than the standard CB model, making the proposed PtC framework better suited for demand forecasting under data scarcity. 5 PtC in Inventory Management System 5.1 Model Formulation in Inventory Management System The forecasting experiments show that PtC improves prediction accuracy, but the operational value of a forecast depends on how it affects downstream decisions. We therefore embed the corrected forecasts into a periodic-review inventory management system. At the beginning of period t, the retailer observes the inventory position IPtIP_t, outstanding replenishment orders, and the current demand context. A replenishment policy then chooses an order quantity qt≥0q_t≥ 0. Orders arrive after a deterministic lead time L, and demand yty_t is realized at the end of the period. Let ItI_t denote the net inventory after demand is realized, where positive values represent inventory and negative values represent unmet demand. The inventory transition is written as: It+1=It+qt−L−yt,I_t+1=I_t+q_t-L-y_t, (23) where qt−Lq_t-L is the order placed L periods earlier. The single-period inventory cost is: Ct=h[It+1]++b[−It+1]+,C_t=h[I_t+1]^++b[-I_t+1]^+, (24) where h is the unit holding-cost coefficient, b is the unit shortage-cost coefficient, and [x]+=maxx,0[x]^+= \x,0\. The objective is to minimize the average cost over the evaluation horizon: C¯=1T∑t=1TCt. C= 1T _t=1^TC_t. (25) PtC enters this system through the demand signal used for replenishment. The baseline ML forecast y^tML y_t^ML is first corrected by the learned contextual-bandit action ata_t, producing: y^tPtC=y^tML(1+at). y_t^PtC= y_t^ML(1+a_t). (26) The corrected forecast is then used to estimate lead-time demand. For a lead time L, the PtC-based replenishment target can be expressed as: D^t,LPtC=∑j=0L−1y^t+jPtC, D_t,L^PtC= _j=0^L-1 y_t+j^PtC, (27) and the corresponding order quantity is: qtPtC=max0,D^t,LPtC−IPt.q_t^PtC= \0, D_t,L^PtC-IP_t\. (28) This formulation keeps the inventory decision simple while allowing the order quantity to respond to online forecast corrections. 5.2 Baseline Policies and Experiment Settings We compare PtC with three inventory-control baselines: base-stock (BS), proximal policy optimization (PPO), and soft actor-critic (SAC). BS is a classical order-up-to policy that replenishes inventory toward a target level based on the estimated lead-time demand. It provides a transparent operational benchmark for forecast-driven inventory control. PPO and SAC represent deep reinforcement learning policies for inventory management, a setting in which the policy learns order decisions from repeated interaction with the inventory simulator (Boute et al., 2022; Gijsbrechts et al., 2022). PPO is an on-policy actor-critic algorithm that updates the policy through a clipped surrogate objective, which improves training stability by limiting overly large policy updates. SAC is an off-policy actor-critic algorithm that maximizes both expected return and policy entropy, encouraging exploration in continuous-control problems (Haarnoja et al., 2018). In this experiment, both RL baselines observe the same inventory state information and are trained to minimize the cumulative inventory cost. All policies are evaluated under the same demand sequences, cost parameters, and lead-time settings. We vary the lead time from L=2L=2 to L=4L=4 to test whether the policies remain robust when the replenishment decision must anticipate demand further into the future. Inventory cost is reported in millions, and lower values indicate better downstream operational performance. 5.3 Numerical Results Figure 5 reports the inventory-cost comparison across the tested lead-time settings. PtC achieves the lowest inventory cost across the reported settings, indicating that the forecast improvements observed in the earlier experiments translate into downstream cost reductions. The result is important because a lower forecast error does not automatically guarantee a lower inventory cost; the correction must also improve the timing and order quantities of replenishment under the uncertainty of lead time. Figure 5: Inventory-cost comparison of BS, PPO, SAC and PtC under different lead times. The comparison also shows different failure modes across baselines. PPO yields the highest inventory costs in the reported settings, suggesting that the learned policy is less stable for this demand-correction task. SAC performs better than PPO but remains above PtC, indicating that a generic continuous-control policy may still struggle when the main uncertainty comes from demand forecasting errors. BS is competitive in some settings because it directly links replenishment to estimated lead-time demand, but it does not adapt the forecast itself when real-time demand feedback reveals bias. PtC combines these two advantages: it preserves a simple forecast-driven replenishment structure and improves the input forecast through online correction. As the lead time increases, inventory decisions become more sensitive to forecast bias because orders must cover a longer demand interval. PtC maintains the lowest cost under the tested lead times, suggesting that correcting the forecast before the replenishment decision can reduce both excess inventory and shortage risk. These results provide downstream evidence for the practical value of the PtC loop in inventory management systems. 6 Conclusion This paper studies adaptive demand forecasting in retail supply chains, where replenishment, pricing, purchasing, and inventory decisions require timely forecasts but demand patterns can shift before static forecasting models are fully retrained. To address this problem, we propose a PtC framework that preserves the base forecast generated by an offline machine-learning model and then applies a few-shot CB correction policy to adjust the forecasts as new feedback arrives. The main finding is that forecast correction can be treated as a sequential learning problem: instead of replacing the original forecaster, a lightweight correction layer can improve predictive reliability and downstream decision quality while retaining the structure learned from historical demand. This study contributes to the forecasting and retail operations literature by filling a gap between static forecasting and full model retraining. Existing forecasting models often produce predictions as fixed outputs, while many operational decision models assume that demand information is already available in a usable form. PtC connects these two views by modeling the post-forecast stage as an adaptive correction process, thereby making demand forecasting part of a continuous learning loop. The similar-SKU augmentation and top-p masked update further clarify how a correction policy can balance responsiveness to new observations with the preservation of previously learned demand regularities. In practice, this design can be embedded into existing forecast pipelines with limited disruption, helping retailers update forecasts for early-cycle, sparse, low-volume, or intermittent items without rebuilding the entire forecasting system. The strength of this study lies in its modular design and its direct connection between forecast accuracy and operational usefulness. Because PtC is placed after the base forecaster, it can work as an adaptive layer rather than as a replacement for existing forecasting models, which makes the framework easier to interpret and deploy. At the same time, the current study has several limitations. The correction policy is built at the SKU level and does not yet explicitly model substitution, complementarity, shelf-space competition, or capacity coupling across products. The downstream decision setting is also simplified, and the robustness of the correction range and top-p update needs further examination under more diverse retail conditions. Future research can extend this work in several directions. First, multi-product correction policies can be developed by incorporating graph, hierarchical, or attention-based structures to capture dependencies across SKUs. Second, forecast correction can be integrated more closely with replenishment, pricing, allocation, and purchasing decisions so that prediction and decision optimization are learned in a more unified loop. Third, future studies should examine online deployment issues such as delayed feedback, changing assortment structures, promotion shocks, stochastic lead times, and service-level constraints. These extensions would move PtC from adaptive forecast correction toward a broader closed-loop decision framework for retail supply chains. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References A. Aamer, L. Eka Yani, and I. Alan Priyatna (2021) Data analytics in the supply chain management: review of machine learning applications in demand forecasting. Operations and Supply Chain Management: An International Journal 14 (1), p. 1–13. Cited by: §2.2. L. Aburto and R. Weber (2007) Improved supply chain management based on hybrid demand forecasts. Applied Soft Computing 7 (1), p. 136–144. Cited by: §2.2. M. Adya and F. Collopy (1998) How effective are neural networks at forecasting and prediction? a review and evaluation. Journal of Forecasting 17 (5-6), p. 481–495. Cited by: §2.2. M. Alley, M. Biggs, R. Hariss, C. Herrmann, M. L. Li, and G. Perakis (2023) Pricing for heterogeneous products: analytics for ticket reselling. Manufacturing & Service Operations Management 25 (2), p. 409–426. Cited by: §2.2. I. Alon, M. Qi, and R. J. Sadowski (2001) Forecasting aggregate retail sales:: a comparison of artificial neural networks and traditional methods. Journal of Retailing and Consumer Services 8 (3), p. 147–156. Cited by: §2.2. R. L. Andrews, I. S. Currim, P. Leeflang, and J. Lim (2008) Estimating the scan⁎ pro model of store sales: hb, fm or just ols?. International Journal of Research in Marketing 25 (1), p. 22–33. Cited by: §2.1. N. S. Arunraj and D. Ahrens (2015) A hybrid seasonal autoregressive integrated moving average and quantile regression for daily food sales forecasting. International Journal of Production Economics 170, p. 321–335. Cited by: §2.1. L. Bethune, D. Grangier, D. Busbridge, E. Gualdoni, M. Cuturi, and P. Ablin (2025) Scaling laws for forgetting during finetuning with pretraining data injection. arXiv Preprint arXiv:2502.06042. Cited by: §2.2, §3.2.4. T. Boone, R. Ganeshan, R. L. Hicks, and N. R. Sanders (2018) Can Google Trends improve your sales forecast?. Production and Operations Management 27 (10), p. 1770–1774. Cited by: §2.1, §2.2. R. N. Boute, S. M. Disney, M. R. Lambrecht, B. Mantin, and J. Gijsbrechts (2022) Deep reinforcement learning for inventory control: a roadmap. European Journal of Operational Research 298 (2), p. 401–412. Cited by: §2.2, §5.2. F. Caro and J. Gallien (2010) Inventory management of a fast-fashion retail network. Operations Research 58 (2), p. 257–273. Cited by: §2.1. T. Chen and C. Guestrin (2016) Xgboost: a scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, p. 785–794. Cited by: §1, §4.1.3. M. C. Cohen, P. Gras, A. Pentecoste, and R. Zhang (2022) Demand prediction in retail: a practical guide to leverage data and predictive analytics. Springer. Cited by: §1. R. Cui, S. Gallino, A. Moreno, and D. J. Zhang (2018) The operational value of social media information. Production and Operations Management 27 (10), p. 1749–1769. Cited by: §2.1, §2.2. B. Fahimnia, T. Tan, and N. Tahirov (2025) Service-level anchoring in demand forecasting: the moderating impact of retail promotions and product perishability. International Journal of Forecasting 41 (2), p. 554–570. Cited by: §1. J. Feizabadi (2022) Machine learning demand forecasting and supply chain performance. International Journal of Logistics Research and Applications 25 (2), p. 119–142. Cited by: §2.2. K. J. Ferreira, B. H. A. Lee, and D. Simchi-Levi (2016) Analytics for an online retailer: demand forecasting and price optimization. Manufacturing & Service Operations Management 18 (1), p. 69–88. Cited by: §2.1, §2.2. R. Fildes, S. Ma, and S. Kolassa (2022) Retail forecasting: research and practice. International Journal of Forecasting 38 (4), p. 1283–1318. Cited by: §2.1. R. Fildes, K. Nikolopoulos, S. F. Crone, and A. Syntetos (2008) Forecasting and operational research: a review. Journal of the Operational Research Society 59 (9), p. 1150–1172. Cited by: §2.1. J. H. Friedman (2001) Greedy function approximation: a gradient boosting machine. The Annals of Statistics 29 (5), p. 1189–1232. Cited by: §1. C. Gao, N. Zhang, Y. Li, Y. Lin, and H. Wan (2023) Adversarial self-attentive time-variant neural networks for multi-step time series forecasting. Expert Systems with Applications 231, p. 120722. Cited by: §2.2. E. S. Gardner Jr (1985) Exponential smoothing: the state of the art. Journal of Forecasting 4 (1), p. 1–28. Cited by: §4.1.3. M. D. Geurts and J. P. Kelly (1986) Forecasting retail sales using alternative models. International Journal of Forecasting 2 (3), p. 261–272. Cited by: §2.1. J. Gijsbrechts, R. N. Boute, J. A. Van Mieghem, and D. J. Zhang (2022) Can deep reinforcement learning improve inventory management? performance on lost sales, dual-sourcing, and multi-echelon problems. Manufacturing & Service Operations Management 24 (3), p. 1349–1368. Cited by: §2.2, §5.2. A. Graves and J. Schmidhuber (2005) Framewise phoneme classification with bidirectional lstm and other neural network architectures. Neural Networks 18 (5-6), p. 602–610. Cited by: §2.2. K. C. Green and J. S. Armstrong (2015) Simple versus complex forecasting: the evidence. Journal of Business Research 68 (8), p. 1678–1685. Cited by: §2.1. X. Guo, K. C. Lichtendahl, and Y. Grushka-Cockayne (2025) Bayesian ensembles of exponentially smoothed life-cycle forecasts. Manufacturing & Service Operations Management 27 (1), p. 230–248. Cited by: §2.2. T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine (2018) Soft actor-critic: off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International conference on machine learning, p. 1861–1870. Cited by: §5.2. S. Hochreiter and J. Schmidhuber (1997) Long short-term memory. Neural Computation 9 (8), p. 1735–1780. Cited by: §1. T. Huang, R. Fildes, and D. Soopramanien (2014) The value of competitive information in forecasting fmcg retail product sales and the variable selection problem. European Journal of Operational Research 237 (2), p. 738–748. Cited by: §2.1. T. Huang, R. Fildes, and D. Soopramanien (2019) Forecasting retailer product sales in the presence of structural change. European Journal of Operational Research 279 (2), p. 459–470. Cited by: §1, §2.1. R. J. Hyndman and G. Athanasopoulos (2018) Forecasting: principles and practice. OTexts. Cited by: §2.1. C. Ingle, D. Bakliwal, J. Jain, P. Singh, P. Kale, and V. Chhajed (2021) Demand forecasting: literature review on various methodologies. In 2021 12th International Conference on Computing Communication and Networking Technologies, p. 1–7. Cited by: §2.2. S. Islam, S. H. Amin, and L. J. Wardley (2021) Machine learning and optimization models for supplier selection and order allocation planning. International Journal of Production Economics 242, p. 108315. Cited by: §2.2. B. Jin and X. Xu (2025a) A study of contemporaneous residential real estate price causation across major Jiangsu province cities: methodology using vector error-correction models and directed acyclic graphs. Economics Open, p. 2550008. Cited by: §2.2. B. Jin and X. Xu (2025b) High-frequency CSI300 spot and futures price predictions via the neural network. Journal of Uncertain Systems, p. 2550008. Cited by: §2.2. G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T. Liu (2017) Lightgbm: a highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems 30. Cited by: §4.1.3. M. Khouja and J. Zhou (2019) Early sale of seasonal inventory in the newsvendor problem. IISE Transactions 51 (6), p. 672–689. Cited by: §1. B. Kim, J. G. Kim, and S. Lee (2024) A multi-agent reinforcement learning model for inventory transshipments under supply chain disruption. IISE Transactions 56 (7), p. 715–728. Cited by: §1. J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. (2017) Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114 (13), p. 3521–3526. Cited by: §2.2, §3.2.4. D. Lei, Y. Qi, S. Liu, D. Geng, J. Zhang, H. Hu, and Z. M. Shen (2024) Pooling and boosting for demand prediction in retail: a transfer learning approach. Manufacturing & Service Operations Management. Cited by: §1. D. Lei, H. Hu, D. Geng, J. Zhang, Y. Qi, S. Liu, and Z. M. Shen (2023) New product life cycle curve modeling and forecasting with product attributes and promotion: a Bayesian functional approach. Production and Operations Management 32 (2), p. 655–673. Cited by: §2.2. S. Li, Y. He, H. Huang, J. Lin, and D. Ivanov (2024a) Supply chain hoarding and contingent sourcing strategies in anticipation of price hikes and product shortages. IISE Transactions 56 (7), p. 746–761. Cited by: §1. X. Li, Y. Xu, R. Law, and S. Wang (2024b) Enhancing tourism demand forecasting with a transformer-based framework. Annals of Tourism Research 107, p. 103791. Cited by: §1, §2.2. B. J. Ma, I. Jackson, M. Huang, S. Villegas, and J. Macias-Aguayo (2025) A data-driven and context-aware approach for demand forecasting in the beverage industry. International Journal of Logistics Research and Applications, p. 1–28. Cited by: §3.1. S. Ma, R. Fildes, and T. Huang (2016) Demand forecasting with high dimensional data: the case of sku retail sales forecasting with intra-and inter-category promotional information. European Journal of Operational Research 249 (1), p. 245–257. Cited by: §1, §2.1. S. Makridakis, A. Andersen, R. Carbone, R. Fildes, M. Hibon, R. Lewandowski, J. Newton, E. Parzen, and R. Winkler (1982) The accuracy of extrapolation (time series) methods: results of a forecasting competition. Journal of Forecasting 1 (2), p. 111–153. Cited by: §2.1. S. Makridakis, M. Hibon, E. Lusk, and M. Belhadjali (1993) The M2-competition: a real-time judgmentally based forecasting study. International Journal of Forecasting 9 (1), p. 5–22. Cited by: §2.1. S. Makridakis and M. Hibon (2000) The M3-competition: results, conclusions and implications. International Journal of Forecasting 16 (4), p. 451–476. Cited by: §2.1. S. Makridakis, E. Spiliotis, and V. Assimakopoulos (2022) M5 accuracy competition: results, findings, and conclusions. International Journal of Forecasting 38 (4), p. 1346–1364. Cited by: §1, §2.2, §4.1.1. P. Malviya and V. Bhandari (2024) A systematic study on effective demand prediction using machine learning. Journal of Integrated Science and Technology 12 (1), p. 711. Cited by: §2.2. M. Mamdouh, M. Ezzat, and H. Hefny (2024) Improving flight delays prediction by developing attention-based bidirectional lstm network. Expert Systems with Applications 238, p. 121747. Cited by: §2.2. R. Mandania and F. S. Oliveira (2023) Dynamic pricing of regulated field services using reinforcement learning. IISE Transactions 55 (10), p. 1022–1034. Cited by: §1. P. Mazumder, P. Singh, and P. Rai (2021) Few-shot lifelong learning. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35, p. 2337–2345. Cited by: §1, §2.2. McKinsey (2022) Stronger forecasting in operations management—even with weak data. Note: Accessed December 19, 2025 External Links: Link Cited by: §1. K. Nikolopoulos, A. A. Syntetos, J. E. Boylan, F. Petropoulos, and V. Assimakopoulos (2011) An aggregate-disaggregate intermittent demand approach (ADIDA) to forecasting: an empirical proposition and analysis. Journal of the Operational Research Society 62 (3), p. 544–554. Cited by: §2.1. F. Petropoulos, D. Apiletti, V. Assimakopoulos, M. Z. Babai, D. K. Barrow, S. B. Taieb, C. Bergmeir, R. J. Bessa, J. Bijak, J. E. Boylan, et al. (2022) Forecasting: theory and practice. International Journal of Forecasting 38 (3), p. 705–871. Cited by: §1, §2.2. S. Punia and S. Shankar (2022) Predictive analytics for demand forecasting: a deep learning-based decision support system. Knowledge-Based Systems 258, p. 109956. Cited by: §1. S. Steinker, K. Hoberg, and U. W. Thonemann (2017) The value of weather information for e-commerce operations. Production and Operations Management 26 (10), p. 1854–1874. Cited by: §2.1, §2.2. X. Tao, X. Hong, X. Chang, S. Dong, X. Wei, and Y. Gong (2020) Few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 12183–12192. Cited by: §3.2.4. J. W. Taylor (2003) Exponential smoothing with a damped multiplicative trend. International Journal of Forecasting 19 (4), p. 715–725. Cited by: §2.1. J. W. Taylor (2007) Forecasting daily supermarket sales using exponentially weighted quantile regression. European Journal of Operational Research 178 (1), p. 154–167. Cited by: §2.1. S. Tian, L. Li, W. Li, H. Ran, X. Ning, and P. Tiwari (2024) A survey on few-shot class-incremental learning. Neural Networks 169, p. 307–324. Cited by: §2.2, §3.2.4. J. R. Trapero, N. Kourentzes, and R. Fildes (2015) On the identification of sales forecasting models in the presence of promotions. Journal of the Operational Research Society 66 (2), p. 299–307. Cited by: §2.1. J. R. Trapero, D. J. Pedregal, R. Fildes, and N. Kourentzes (2013) Analysis of judgmental adjustments in the presence of promotions. International Journal of Forecasting 29 (2), p. 234–243. Cited by: §2.1. M. R. Wagner (2015) Robust purchasing and information asymmetry in supply chains with a price-only contract. IISE Transactions 47 (8), p. 819–840. Cited by: §1. X. Wang, H. Wang, S. Li, and H. Jin (2024) A reinforcement learning-based online learning strategy for real-time short-term load forecasting. Energy 305, p. 132344. Cited by: §1, §2.2. X. Xu and Y. Zhang (2021) Corn cash price forecasting with neural networks. Computers and Electronics in Agriculture 184, p. 106120. Cited by: §2.2. X. Xu and Y. Zhang (2023) Price forecasts of ten steel products using Gaussian process regressions. Engineering Applications of Artificial Intelligence 126, p. 106870. Cited by: §2.2. L. Ye, N. Xie, J. E. Boylan, and Z. Shang (2024) Forecasting seasonal demand for retail: a fourier time-varying grey model. International Journal of Forecasting 40 (4), p. 1467–1485. Cited by: §2.1. G. P. Zhang, Y. Xia, and M. Xie (2024) Intermittent demand forecasting with transformer neural networks. Annals of Operations Research 339 (1), p. 1051–1072. Cited by: §1, §2.2. G. Zhang, L. Wang, G. Kang, L. Chen, and Y. Wei (2023) Slca: slow learner with classifier alignment for continual learning on a pre-trained model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 19148–19158. Cited by: §3.2.4. W. Zhang, Y. Yu, Y. Qi, F. Shu, and Y. Wang (2019) Short-term traffic flow prediction based on spatio-temporal analysis and cnn deep learning. Transportmetrica A: Transport Science 15 (2), p. 1688–1711. Cited by: §2.2. F. Zhong, Z. Zhou, and M. Leng (2020) Game-theoretic analyses of strategic pricing decision problems in supply chains. IISE Transactions 53 (6), p. 704–718. Cited by: §1. Supplemental Online Materials to “A Predict-then-Correct Loop Based on Few-Shot Continuous Contextual Bandit for Demand Forecasting” Appendix A: Dataset Splitting Table A1: Dataset and splitting Dataset Cluster Quantity Training Validation Forecast Sliding window Walmart Stable & High Volume 217 Week 1–134 Week 135–143 Week 144–152 3 Stable & Low Volume 3122 Week 1–134 Week 135–143 Week 144–152 3 Erratic & Intermittent 3256 Week 1–134 Week 135–143 Week 144–152 3 Beverage Stable & High Volume 52 Day 1–1765 Day 1766–1825 Day 1826–1885 3 Stable & Low Volume 465 Day 1–1765 Day 1766–1825 Day 1826–1885 3 Erratic & Intermittent 168 Day 1–1765 Day 1766–1825 Day 1826–1885 3 Appendix B: Principal Hyperparameter Settings Table B1: Principal hyperparameter settings Setting Value Batch size 64 Learning rate 3×10−43× 10^-4 Optimizer Adam Hidden dimension 256 Action dimension 1 Observation dimension 15 Entropy coefficient α 0.02 Network layers 3 Parameter update ratio for the Walmart dataset 0.10 Parameter update ratio for the beverage dataset 0.16 Sliding-window length 3 Appendix C: Augmentation Data Selection Figure C1: Augmentation data selection within each cluster. Appendix D: Proof of Proposition 1 Proof of Proposition 1. Since 0∈0 and π0(st)=0 _0(s_t)=0 belongs to the policy class Π , the PtC framework can choose the zero-correction policy. Under this policy, the corrected forecast becomes y^tPtC(π0)=y^tML(1+π0(st))=y^tML. y_t^PtC( _0)= y_t^ML(1+ _0(s_t))= y_t^ML. Therefore, the ML-only prediction is a special case of the PtC prediction class. Since the PtC framework optimizes over a feasible policy class that contains the zero-correction policy, taking the infimum over Π cannot yield a larger expected loss than the loss attained by π0 _0. Hence, infπ∈Π[1T∑t=1Tℓ(yt,y^tML(1+π(st)))]≤[1T∑t=1Tℓ(yt,y^tML(1+π0(st)))]. _π∈ E [ 1T _t=1^T (y_t, y_t^ML(1+π(s_t)) ) ] [ 1T _t=1^T (y_t, y_t^ML(1+ _0(s_t)) ) ]. Using π0(st)=0 _0(s_t)=0, the right-hand side reduces to [1T∑t=1Tℓ(yt,y^tML)].E [ 1T _t=1^T (y_t, y_t^ML ) ]. ∎ Appendix E: Proof of Proposition 2 Proof of Proposition 2. Let ℒnew(θ)L_new(θ) denote the loss on the new demand cycle, and define the feasible parameter space under top-p updating as Θp=θ0+Δ:supp(Δ)⊆Sp. _p= \θ^0+ :supp( ) S_p \. Then the adaptation gap Gad(p)=infθ∈Θpℒnew(θ)−infθ∈Θ1ℒnew(θ)G_ad(p)= _θ∈ _pL_new(θ)- _θ∈ _1L_new(θ) is monotone non-increasing with respect to p, and satisfies Gad(1)=0G_ad(1)=0. Moreover, suppose that the masked gradient update is performed for U steps: θpu+1=θpu−ηmp⊙gpu,u=0,…,U−1,θ^u+1_p=θ^u_p-η\,m_p g^u_p, u=0,…,U-1, where θp0=θ0θ^0_p=θ^0, η>0η>0 is the learning rate, and every gradient coordinate is bounded as |(gpu)j|≤G|(g^u_p)_j|≤ G. Then the parameter drift satisfies ‖θpU−θ0‖2≤ηUG⌈pd⌉.\|θ^U_p-θ^0\|_2≤η UG pd . Therefore, the upper bound on parameter drift is monotone non-decreasing with respect to p and equals zero when p=0p=0. If the old-cycle loss ℒold(θ)L_old(θ) is LoldL_old-Lipschitz continuous, then ℒold(θpU)−ℒold(θ0)≤LoldηUG⌈pd⌉,L_old(θ^U_p)-L_old(θ^0)≤ L_oldη UG pd , which implies that the forgetting-risk bound is also monotone non-decreasing with respect to p. We first prove the plasticity result. For any p1≤p2p_1≤ p_2, the nesting property of the Top-p selected sets gives Sp1⊆Sp2S_p_1 S_p_2. Therefore, any parameter perturbation supported on Sp1S_p_1 is also supported on Sp2S_p_2, which implies Θp1⊆Θp2. _p_1 _p_2. Since Θp2 _p_2 is a larger feasible parameter space, minimizing the new-cycle loss over Θp2 _p_2 cannot yield a larger optimal value than minimizing over Θp1 _p_1. Hence, infθ∈Θp2ℒnew(θ)≤infθ∈Θp1ℒnew(θ). _θ∈ _p_2L_new(θ)≤ _θ∈ _p_1L_new(θ). Subtracting the constant full-update benchmark infθ∈Θ1ℒnew(θ) _θ∈ _1L_new(θ) from both sides gives Gad(p2)≤Gad(p1).G_ad(p_2)≤ G_ad(p_1). Thus, the adaptation gap is monotone nonincreasing in p. When p=1p=1, all parameters are allowed to update, so Θp=Θ1 _p= _1 and therefore Gad(1)=0.G_ad(1)=0. We next prove the stability result. By the masked update rule, θpU−θ0=−η∑u=0U−1mp⊙gpu.θ^U_p-θ^0=-η _u=0^U-1m_p g^u_p. Taking the Euclidean norm and applying the triangle inequality yields ‖θpU−θ0‖2≤η∑u=0U−1‖mp⊙gpu‖2.\|θ^U_p-θ^0\|_2≤η _u=0^U-1\|m_p g^u_p\|_2. Since the mask mpm_p contains exactly ⌈pd⌉ pd nonzero entries and each gradient coordinate is bounded by G, we have ‖mp⊙gpu‖2≤G⌈pd⌉.\|m_p g^u_p\|_2≤ G pd . Therefore, ‖θpU−θ0‖2≤ηUG⌈pd⌉.\|θ^U_p-θ^0\|_2≤η UG pd . Because ⌈pd⌉ pd is monotone non-decreasing in p, the parameter-drift upper bound is also monotone non-decreasing in p. When p=0p=0, no parameter is selected for updating, so mp=0m_p=0 and ‖θ0U−θ0‖2=0.\|θ^U_0-θ^0\|_2=0. Finally, if ℒoldL_old is LoldL_old-Lipchitz continuous, then ℒold(θpU)−ℒold(θ0)≤Lold‖θpU−θ0‖2.L_old(θ^U_p)-L_old(θ^0)≤ L_old\|θ^U_p-θ^0\|_2. Substituting the drift bound gives ℒold(θpU)−ℒold(θ0)≤LoldηUG⌈pd⌉.L_old(θ^U_p)-L_old(θ^0)≤ L_oldη UG pd . Thus, increasing p expands the adaptable parameter space and improves plasticity, but it also increases the upper bound on parameter drift and forgetting risk. This establishes the stability-plasticity trade-off of the top-p updating rule. ∎ Appendix F: Proof of Proposition 3 Proof of Proposition 3. Define the smoothed expected correction objective ϵ(θ)=(st,yt,y^tML)∼,at∼πθ[ρϵ(y^tML(1+at)−yt)],J_ε(θ)=E_(s_t,y_t, y_t^ML) ,\,a_t _θ [ _ε ( y_t^ML(1+a_t)-y_t ) ], where ρϵ(z)=z2+ϵ2 _ε(z)= z^2+ε^2. The smoothing parameter ϵ>0ε>0 avoids the non-differentiability of the absolute value while preserving the MAE-type correction objective as ϵ→0ε→ 0. The notation at∼πθa_t _θ means that the correction action at time t is generated by the policy parameterized by θ. At update step u, the top-p rule defines a binary mask mu∈0,1dm_u∈\0,1\^d. The masked policy update can be written as θu+1=θu−ηmu⊙∇ϵ(θu).θ^u+1=θ^u-η\,m_u _ε(θ^u). Let du=mu⊙∇ϵ(θu).d_u=m_u _ε(θ^u). Then θu+1−θu=−ηduθ^u+1-θ^u=-η d_u. Since ϵJ_ε is L-smooth, the descent lemma gives ϵ(θu+1)≤ϵ(θu)+∇ϵ(θu)⊤(θu+1−θu)+L2‖θu+1−θu‖22.J_ε(θ^u+1) _ε(θ^u)+ _ε(θ^u) (θ^u+1-θ^u)+ L2\|θ^u+1-θ^u\|_2^2. Substituting θu+1−θu=−ηduθ^u+1-θ^u=-η d_u yields ϵ(θu+1)≤ϵ(θu)−η∇ϵ(θu)⊤du+Lη22‖du‖22.J_ε(θ^u+1) _ε(θ^u)-η _ε(θ^u) d_u+ Lη^22\|d_u\|_2^2. Because mum_u is a binary mask, ∇ϵ(θu)⊤du=∇ϵ(θu)⊤(mu⊙∇ϵ(θu))=‖du‖22. _ε(θ^u) d_u= _ε(θ^u) (m_u _ε(θ^u) )=\|d_u\|_2^2. Therefore ϵ(θu+1)≤ϵ(θu)−η(1−Lη2)‖du‖22.J_ε(θ^u+1) _ε(θ^u)-η (1- Lη2 )\|d_u\|_2^2. If 0<η≤1/L0<η≤ 1/L, then 1−Lη/2≥1/21-Lη/2≥ 1/2, and thus ϵ(θu+1)≤ϵ(θu)−η2‖du‖22≤ϵ(θu).J_ε(θ^u+1) _ε(θ^u)- η2\|d_u\|_2^2 _ε(θ^u). This proves the monotonic decrease of the expected correction loss after each policy update. ∎ Appendix G: Top-p Masked Bandit Update Algorithm G1 Top-p masked bandit update for the CB correction policy 1: Input: Feedback data newD_new, initial policy πθ0 _θ^0, update ratio p, learning rate η, mini-batch size B, and update steps U 2: Output: Updated correction policy πθU _θ^U 3: for u=0,…,U−1u=0,…,U-1 do 4: Sample a mini-batch ℬu⊂newB_u _new, where |ℬu|=B|B_u|=B 5: For all t∈ℬut _u, compute μθu(st) _θ^u(s_t) and σθu(st) _θ^u(s_t), sample a~t∼(μθu(st),σθu2(st)), a_t ( _θ^u(s_t), _θ^u^2(s_t) ), and set at=clip(a~t,−1,2)a_t=clip( a_t,-1,2) 6: Compute the corrected forecasts and rewards: rt=τ(|y^tML−yt|−|y^tCB−yt|)r_t=τ (| y^ML_t-y_t|-| y^CB_t-y_t| ) 7: Estimate the bandit policy gradient: ∇^θJ(θu)=1|ℬu|∑t∈ℬu(rt−bt)∇θlogπθu(at∣st) ∇_θJ(θ^u)= 1|B_u| _t _u(r_t-b_t) _θ _θ^u(a_t s_t) 8: For each layer k, construct mask M by selecting the smallest qk=⌈p|θk|⌉q_k= p|θ^k| parameters according to |θjk|| _j^k| 9: Update: θu+1=θu+ηM⊙∇^θJ(θu)θ^u+1=θ^u+η M ∇_θJ(θ^u) 10: end for 11: Return πθU _θ^U