Introduction

  • TL;DR: AI Sales Forecasting must stay consistent across planning levels (total/category/SKU). The common production pattern is (1) generate base forecasts, then (2) apply forecast reconciliation (e.g., MinT) to enforce coherence. For new items, “cold-start” is solved by borrowing signal from hierarchies and similar items (metadata/content/price tiers). Promotions should be designed either as model features or as a separate uplift (counterfactual) estimation pipeline (e.g., CausalImpact/BSTS).

Why it matters: Without coherence, different teams will operate on different numbers, breaking replenishment and planning alignment.

Prerequisites

  • Fixed hierarchy mappings (SKU→brand→category→total, store→region→total) and unique keys
  • Promotion schema (start/end, discount depth, mechanics) and data quality gates
  • Label slices: promo vs non-promo, launch window for new items

Why it matters: Most failures come from inconsistent definitions and broken joins, not model choice.

1) Hierarchical forecasting with reconciliation

Base forecasts → Reconciled (coherent) forecasts

  • The standard framework is described in FPP3 with summing-matrix notation.
  • MinT is a widely cited “optimal” reconciliation approach.
  • Practical tooling: sktime ReconcilerForecaster and Nixtla HierarchicalForecast document reconciliation support.

Why it matters: Coherence is a hard constraint for planning systems, not a “nice-to-have.”

2) Cold-start new items

  • Borrow signal from hierarchies + pool information from similar items using metadata and content features.
  • Watch for censored sales due to stockouts/low exposure early in the lifecycle.

Why it matters: Cold-start error is often dominated by data generation (availability/exposure), not forecasting math.

3) Promotion uplift design

Two patterns:

  • Feature-based forecasting: include promotion variables; classic work shows explicit promo features can improve promo-period accuracy.
  • Uplift as causal/counterfactual estimation: BSTS/CausalImpact estimates “what would have happened without the promo.”

Why it matters: Promo periods can induce systematic under-forecasting; monitor promo slices explicitly.

Conclusion

  • Use reconciliation to enforce coherent totals across the hierarchy.
  • Solve cold-start by pooling and similarity-based seeding, then retrain quickly.
  • Choose promo modeling intentionally: features vs separate uplift estimation.

Summary

  • Coherence requires reconciliation, not wishful thinking.
  • Cold-start = borrow signal from hierarchy + similar items.
  • Promotions need dedicated evaluation slices and, when needed, counterfactual uplift pipelines.

#AISalesForecasting #HierarchicalForecasting #ForecastReconciliation #MinT #ColdStart #PromotionUplift #CausalImpact #RetailAnalytics

References

  • (Forecast reconciliation, 2026-02-10)[https://otexts.com/fpp3/reconciliation.html]
  • (Optimal forecast reconciliation for hierarchical and grouped time series (MinT), 2017)[https://robjhyndman.com/papers/mint.pdf]
  • (ReconcilerForecaster API, 2026-02-10)[https://www.sktime.net/en/stable/api_reference/auto_generated/sktime.forecasting.reconcile.ReconcilerForecaster.html]
  • (HierarchicalForecast documentation, 2026-02-10)[https://nixtlaverse.nixtla.io/hierarchicalforecast/index.html]
  • (Deep learning for new fashion product demand prediction, 2026)[https://link.springer.com/article/10.1007/s41060-025-00888-8]
  • (SKU demand forecasting in the presence of promotions, 2009)[https://www.sciencedirect.com/science/article/abs/pii/S0957417409004035]
  • (Inferring causal impact using Bayesian structural time-series models, 2015)[https://research.google.com/pubs/archive/41854.pdf]
  • (CausalImpact documentation, 2026-02-10)[https://google.github.io/CausalImpact/CausalImpact.html]
  • (Challenges for Demand Forecasting in Grocery Retail, 2026)[https://arxiv.org/html/2602.04464v1]
  • (Forecast reconciliation: A review, 2023)[https://www.sciencedirect.com/science/article/pii/S0169207023001097]
  • (A new method to forecast intermittent demand in the presence of obsolescence, 2019)[https://www.sciencedirect.com/science/article/abs/pii/S0925527318300562]
  • (Forecasting hierarchical and grouped time series, 2026-02-10)[https://otexts.com/fpppy/nbs/11-hierarchical-forecasting.html]
  • (Dynamic Dual-Phase Forecasting Model for New Product Demand, 2025)[https://www.mdpi.com/2227-7390/13/10/1613]