Auditing World-Model Explanations in Retail Forecasting: Faithfulness, Accuracy and Price Response

Kathik Iyer

In the world of retail, a forecast is more than just a number; it’s a roadmap for ordering stock. When a merchant plans a promotion, they aren't just asking "how much will I sell?" but "how much extra will this promotion add?" This "extra" is known as lift.

In his paper, "Auditing World-Model Explanations in Retail Forecasting," Karthik Iyer reveals a dangerous gap in how we trust these models. He demonstrates that a model’s explanation can be perfectly "faithful" (it accurately describes what the model is doing) while being fundamentally "incorrect" (it gets the real-world cause-and-effect wrong). For a retailer, this distinction is the difference between a profitable promotion and a warehouse full of unsold goods.

TL;DR

Standard "sanity checks" for AI explanations often only prove that an explanation matches the model's internal logic, not that it reflects reality. This paper introduces an auditing framework for retail forecasting that reveals how models can pass these checks while overstating promotional lift by up to 40%, potentially leading to massive overstocking costs.

What Problem Does This Paper Address?

Retailers use "plan-lift" explanations to decompose a forecast into a baseline (what would sell anyway) and lift (the boost from price cuts, displays, or ads). Merchants use these lift estimates to decide how much inventory to buy.

The challenge is that forecasting models are often "black boxes." To trust them, researchers use "faithfulness" checks to ensure the explanation isn't just making things up. However, Iyer argues that faithfulness is not enough. If a model learns a biased relationship—for example, if it thinks a promotion caused a sales spike that was actually just a seasonal trend—it will provide a "faithful" but "incorrect" explanation. The paper seeks to bridge this gap by auditing explanations against known truths and real-world causal benchmarks.

Methods

The author proposes a suite of eight checks (C1 to C8) to audit any model that provides plan-lift explanations. These checks are categorized into four buckets:

  1. Against the model itself: Does the explanation hold up by construction and remain stable?
  2. Against common sense: Does a deeper price cut ever (incorrectly) lead to a lower forecast?
  3. Against what sold: Does the predicted lift correlate with actual sales?
  4. Against another model and a cause: Does the model agree with established benchmarks (like LightGBM) or measured causal effects?
Figure 1: Each forecast’s explanation is checked against the model itself, common sense, what sold, another model and a measured cause. The explanation is reported in units, baseline plus lift. Pass rules for C1 to C6 are fixed in docs/PAPER2_PLAN.md before any result; C7 and C8 are reported per model (Table 1). Every model version and every check is reported.

The paper evaluates two types of models: a World Model (a neural network that "rolls" through a plan) and LightGBM (a standard gradient-boosted tree model).

Crucially, the author uses a "known-answer" study. By generating synthetic store data where the exact effect of every price cut and promotion is programmed in, the author can definitively measure when a model's explanation is lying.

Datasets

The study utilizes several prominent retail datasets:

  • M5 (Walmart): Daily sales of ~3,000 items.
  • Dominick’s Finer Foods: A classic 1990s dataset from a Chicago grocery chain, used here to estimate real-world price response.
  • Breakfast at the Frat & Corporación Favorita: Public datasets used to test the models across different retail contexts.
  • Synthetic "Known-Answer" Data: A custom-generated dataset where the "ground truth" causal effects are known.

Key Insights and Ideas

  • Faithfulness $\neq$ Correctness: This is the "aha moment." A model can pass every internal sanity check (C1-C3) and still be dangerously wrong about the world. Iyer calls these "faithful failures."
  • The "Merchant Habit" Bias: If merchants historically run promotions during weeks they already expect high demand, models will incorrectly attribute that high demand to the promotion itself. In the synthetic study, this caused models to overstate lift by ~1.4x.
  • The Regular Price Trap: Models often learn from temporary "price cuts" (which have high impact) and incorrectly assume that a permanent "regular price" change will have the same massive effect. In reality, regular price changes move far fewer units.
  • Accuracy is No Shield: Surprisingly, the paper found that a model's forecasting accuracy (how close it gets to the final number) is not related to how many explanation checks it passes. You can have a very accurate forecast built on a completely wrong causal explanation.

Key Results

The results from the "known-answer" study (summarized in Figure 3) show that even "correct" models can be misled by the data they are fed.

  • Lift Overstatement: When plans followed demand (the "high-demand weeks" scenario), the World Model stated 1.33 to 1.40 times the true lift. LightGBM was similarly off (1.26x).
  • Price Response Error: The World Model’s stated response to a lasting regular-price change was 4.7 to 6.5 times the true effect.
  • Common Sense Failures: On the real-world Dominick’s data, a deeper price cut actually lowered the forecast in 29% to 83% of cases—a clear logic error that standard accuracy metrics would never catch.
  • The Cost of Error: Ordering stock based on these overstated lifts would result in an inventory error of 0.44 to 0.48 of the true lift.
Figure 4: When plans follow demand, both the world model and LightGBM overstate what the plan adds, and agree with each other. Each point is an item-week with something planned: the model’s total lift against the true lift, units. Dashed: equal. Top, plans at random; bottom, plans made more often in high-demand weeks.

Limitations and Future Work

The author acknowledges several hurdles:

  • Observational Noise: The real-world benchmark (C7) relies on "zone price deviations" in old data, which is a noisy proxy for a true randomized experiment.
  • Simplicity of Synthetic Data: The generated world doesn't account for "cannibalization" (e.g., buying more Coke because it's on sale, so you buy less Pepsi).
  • Need for Experiments: The paper concludes that for a retailer to truly trust these lifts, they should run their own controlled experiments (A/B tests) to create a "ground truth" benchmark for their specific stores.

In short: Don't just ask if your model is accurate; ask if its explanation of why it's accurate matches the real world.

Summary updated

Abstract

Forecasting models that read the price and promotion plan explain each forecast as a baseline, what would sell with nothing planned, plus the lift the plan adds, and merchants order stock and plan promotions on that lift. We show that an explanation can pass the checks used to trust it and still be wrong enough to cost money. We define eight checks for any forecaster that states a plan lift, six with a pass rule fixed before any result, and run them on a world model and on LightGBM. In generated store data with known effects, the world model passes the checks of construction, stability, price direction and agreement with sales and with LightGBM, yet when merchants plan into high-demand weeks it states 1.33 to 1.40 times the true lift, and 1.23 to 1.42 times in eight variants of that world; LightGBM overstates it too (1.26, and 1.16 to 1.35). Both still pick the right weeks to promote, but stock ordered to the world model's lift is off by 0.44 to 0.48 of the true lift. The world model's stated response to a lasting regular-price change is 4.7 to 6.5 times the true -0.5, an error in kind that only a measured cause finds. Sanity checks for explanations test whether an explanation depends on the model; a decision needs to know whether it is right about the world, and these results show the first does not give the second. For the final tuned world models of our companion paper on nine data sets, baseline and lift add up under each model's own plan on every window, but a deeper price cut lowers the forecast in 29% to 83% of item-week steps on Dominick's. Across 703 configurations from the settings searches, checks passed were not related to accuracy (Spearman 0.02, 95% interval -0.02 to 0.05). A noisy benchmark of price episodes in Dominick's data is reported as a check of direction only.
BibTeX
@misc{iyer2026auditing,
  author = {Kathik Iyer},
  title = {{Auditing World-Model Explanations in Retail Forecasting: Faithfulness, Accuracy and Price Response}},
  year = {2026},
  eprint = {2611.20811},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20811},
}
Figure