When Does Pooling Pay? Credibility and Resolution under Forgetting in Intermittent-Demand Forecasting

arXiv:2511.12749 · stat.ML, cs.LG · Submitted 2025-11-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "When Does Pooling Pay? Credibility and Resolution under Forgetting in Intermittent-Demand Forecasting".

Jane: Intermittent demand forecasting presents significant challenges due to sparse observations and cold-start items, and this paper introduces TSB-HB,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Moving on to the title and the authors, "When Does Pooling Pay? Credibility and Resolution under Forgetting in Intermittent-Demand Forecasting," it immediately makes you wonder about the trade-off between combining data from many items versus sticking strictly to what we know for a single item. Jane, how do you think that tension plays out in the context of this paper's focus?

Jane: Well, the title points directly to that central question: when is it actually beneficial for us to pool information across different items when dealing with intermittent demand? It suggests there are limits to how much we can safely aggregate data without compromising accuracy for individual series.

Lu: The authors tackle this by explicitly modeling the demand occurrence using a Beta-Binomial distribution and nonzero sizes with a Log-Normal distribution, which sets up a very principled way to handle those uncertainties upfront.

Meng: I’m curious about the taxonomy part mentioned in the abstract; they use ADI–CV2 segmentation to group items before applying their model. Does this mean we can automatically classify our inventory items into predictable demand types, like smooth versus lumpy?

Lalam: If it helps us categorize demand patterns, that would be huge for our AI because it allows us to apply different forecasting strategies tailored specifically to the inherent nature of the item's demand behavior.

The paper's summary: Tom: So, summarizing what they actually did in "When Does Pooling Pay? Credibility and Resolution under Forgetting in Intermittent-Demand Forecasting," it seems they’ve developed TSB-HB, a hierarchical Bayesian extension of TSB that uses these distribution models to estimate demand occurrence and size. Jane, can you simplify that generative foundation for our listeners?

Jane: Think of it like this: instead of just guessing the next number based on the past, this method builds a full mathematical story about how demand happens—first modeling when it occurs, and then modeling how large that demand is. This framework is designed to stabilize estimates for items that rarely sell or are brand new by letting information flow between similar items in a controlled way.

Lu: They achieve this by defining occurrence probabilities using a Beta-Binomial distribution and the conditional positive sizes with a Log-Normal distribution, which gives us two distinct ways to model the demand process.

Meng: The paper mentions deriving closed-form item-level posterior moments for both occurrence and size, which is great for speed. But what's the practical reality of fitting those group hyperparameters using empirical Bayes? Is that computationally expensive when we have millions of items?

Lalam: That closed-form derivation combined with low-dimensional empirical Bayes fitting suggests a pathway toward scalable inference because it keeps the complexity manageable while still leveraging that panel information sharing.

The paper's improvements: Tom: The paper clearly outlines several improvements they made, and I want to focus on how they address the weaknesses of older models. Jane, what are the key methodological enhancements they propose in TSB-HB that make it better than just running a standard TSB calculation?

Jane: The main improvement is moving from simple heuristics to a principled generative foundation, which gives us more confidence in our predictions because we have underlying probability distributions rather than just single point estimates. It also introduces hierarchical priors that allow for partial pooling across items, which is key for stabilizing estimates on sparse series.

Lu: They specifically derive closed-form item-level posterior moments for occurrence and size, and they use low-dimensional empirical Bayes hyperparameter fitting, which helps them manage the complexity of the group structure.

Meng: The paper also includes a walk-forward robustness check, which is important because it tests how well this model holds up when we move from in-sample estimation to predicting future data. That kind of validation is what engineers look for before trusting any new forecasting tool.

Lalam: I think the robustness check combined with the way they handle item assignment based on ADI/CV2 taxonomy shows a very thoughtful approach to ensuring that the model works across different types of demand patterns, not just one specific scenario.

Conclusion: Tom: So, wrapping up with the conclusion of "When Does Pooling Pay? Credibility and Resolution under Forgetting in Intermittent-Demand Forecasting," it sounds like they’ve established a method that balances the need for individual item accuracy with the stability gained from sharing information across items. Jane, what's your final thought on the broader implications for our field?

Jane: I think the implication is that we can stop treating every single item in isolation when we know how to intelligently pool information, leading to more reliable forecasts overall and better management of inventory uncertainty. It gives us a much clearer picture of where pooling actually delivers value.

Lu: From a theoretical standpoint, the work provides a coherent generative reinterpretation of classical TSB structures while offering principled panel-wise information sharing, which opens up new avenues for modeling complex time series data.

Meng: For practical application, if this works as well on both the Online Retail and M5 datasets, it suggests we can deploy these kinds of models efficiently in a wide variety of real-world retail scenarios without needing custom tuning for every single product category.

Lalam: Ultimately, the TSB-HB framework provides a more intelligent way for our AI to learn from its experiences across many different items, which will make the entire forecasting ecosystem much more robust and trustworthy.

Zong-Han Bai, Po-Yen Chu

stat.ML, cs.LG

Submitted: 2025-11-16

Updated: 2026-10-02

Code: https://github.com/Nixtla/statsforecast

Importance score: 91/100

The gist: Intermittent demand forecasting presents significant challenges due to sparse observations and cold-start items, and this paper introduces TSB-HB, a hierarchical Bayesian extension that provides a

Key concepts

Intermittent Demand Forecasting
This is the challenge of predicting demand that occurs infrequently and unpredictably, like in retail. The data is sparse, making it hard to reliably estimate future demand patterns without special modeling techniques.
TSB Method
The Teunter–Syntetos–Babai method is a standard approach used to forecast intermittent demand by modeling the occurrence probability and the size of demand separately. It provides a baseline for comparison in this study.
Hierarchical Bayesian Extension (TSB-HB)
This model uses Bayesian statistics with hierarchical priors. It allows information from similar items (groups) to be shared, stabilizing estimates for individual items that have little data, while still allowing each item to retain its unique characteristics.

Terminology

Summary

Intermittent demand forecasting presents significant challenges due to sparse observations and cold-start items, and this paper introduces TSB-HB, a hierarchical Bayesian extension that provides a principled generative foundation by enabling partial pooling across items to stabilize estimates while preserving heterogeneity.

The gist: TSB-HB is a hierarchical Bayesian extension of the Teunter–Syntetos–Babai (TSB) method that models demand occurrence with a Beta–Binomial distribution and nonzero demand sizes with a Log-Normal distribution, using hierarchical priors to enable partial pooling across items to stabilize estimates for sparse or cold-start series while preserving heterogeneity.

Model Formulation and Structure

The TSB-HB model is formulated as a hierarchical, generative analogue of TSB that preserves its multiplicative decomposition while enabling principled panel-wise information sharing. For each item in group g(i), the occurrence probability is modeled as:

πi ∼ Beta(αg(i), βg(i))

The conditional positive-demand size is modeled as:

** li,t (pi,t = 1, µi) ∼ N (µi, σ2g(i))**

The resulting demand is defined multiplicatively as:

(Yi,t = exp(li,t), pi,t = 1; 0; pi,t = 0)

Estimation and Shrinkage Mechanism

The model utilizes empirical-Bayes (EB) estimation to fit group hyperparameters. The core mechanism involves hierarchical shrinkage to stabilize sparse and cold-start series:

  1. Group hyperparameters are fitted using the Beta–Binomial marginal likelihood for occurrence parameters and a random-effects marginal likelihood on item sufficient statistics for size parameters.

  2. The posterior mean for occurrence is given by:

πbi = αg + mi / (αg + βg) + ni / (αg + βg) = λi,g · mi / ni + (1 − λi,g) · µπ, g,

  1. For conditional positive sizes, posterior shrinkage on the log scale uses:

ki = σb2i,proc / τb2, wi = mi / (mi + ki)

Forecasting and Performance

The point forecast remains multiplicative:

  1. The point forecast is defined as:

(Ybi = πbi · Sbi)

  1. The size mean used in the implementation is a plug-in conditional mean:

(Sbi = exp(µbi + 1/2 σb2i,proc))

Probabilistic forecasting yields a zero-inflated Log-Normal predictive law:

  1. For a quantile q, the prediction is defined as:

Qi(q) = 0, if q ≤ 1 − π̂i,

(exp(µ̂i + σi,pred Φ−1 q − (1 − π̂i)) / π̂i, if q > 1 − π̂i)

Empirical Validation and Contributions

The paper demonstrates the effectiveness of TSB-HB through several key findings:

(Online Retail Dataset):

  1. TSB-HB achieves the lowest RMSE (17.6930) and RMSSE (4.7875) among all baselines, representing a 5.5% reduction in RMSE relative to the standard TSB model.

  2. In segment-wise performance analysis, TSB-HB shows marginal improvements in the Intermittent and Lumpy groups, where sparsity and variance make purely local estimation unstable.

  3. Distributional reliability is superior: TSB-HB achieves the lowest Mean Pinball Loss (1.9192) across all models, with a robust 0.8454 coverage at the 80% nominal level, which is preferable in retail context as it offers a safety buffer against unexpected spikes while maintaining sharpness during inactive periods.

(M5 Dataset):

  1. On the M5 sample, TSB-HB achieves the lowest MAE (1.1771) and RMSE (2.8286), outperforming established baselines such as CrostonSBA and ADIDA, suggesting that the regularization benefit is not tied to a single dataset.

(Ablation Study):

  1. Component ablation shows that size stabilization via hierarchical priors is the primary driver of point accuracy, as full HB (p + size) achieves the best performance (MAE 5.7663). Furthermore, Omission of process variance substantially narrows intervals but triggers under-coverage and higher Pinball Loss, proving it essential for reliable predictive distributions.

(Comparison with Deep Learning):

Improvements for AI systems

Based on the provided scientific paper, here are specific improvements that can be made to existing AI forecasting systems, categorized by technical enhancement:

The TSB-HB framework allows for significant upgrades in demand forecasting systems by moving from heuristic smoothing to a principled, generative modeling approach.

  1. Improve robustness of sparse/cold-start item predictions:

  2. Enhance accuracy across different demand patterns (Smooth vs. Intermittent):

  3. Improve distributional reliability and uncertainty quantification:

  4. Increase scalability for large-scale catalogs:

  5. Optimize computational efficiency for high-throughput inventory systems:

Here is what the improved AI system can do based on these enhancements:

The improved AI system, built on the TSB-HB framework, can perform the following specific actions:

  1. Identify demand patterns (Smooth, Erratic, Intermittent, Lumpy) and apply tailored forecasting strategies based on item characteristics (ADI/CV2 taxonomy).

  2. Provide highly accurate point forecasts (lowest RMSE/RMSSE among baselines) for intermittent and lumpy series by leveraging panel-wise information sharing via hierarchical shrinkage toward group priors.

  3. Generate reliable probabilistic forecasts (Pinball Loss, Coverage@80) that accurately capture the long-tail behavior of demand spikes, offering superior calibration compared to standard parametric or conformal methods.

  4. Scale forecasting models efficiently across massive retail catalogs by utilizing closed-form item-level posterior moments and efficient O(N) per-epoch inference suitable for large panels.

  5. Provide stable, scale-aware forecasts under extreme uncertainty (e.g., 100-step horizons), significantly outperforming deep learning alternatives like RNNs in terms of robustness against large forecast excursions, while maintaining sub-millisecond inference times per series.

Related papers