How Deep Meta‑Learning Boosts Spatio‑Temporal Sales Forecasting for Retail

Sales forecasting for online and offline retail is crucial for inventory planning, especially during large shopping festivals, but historical data are scarce.

Background

Spatial distribution of stores, population attributes, and date types (weekday, weekend, festival) lead to diverse shopping behaviors. Limited data per region and time make it hard to train accurate models; heterogeneity of population density and evolving regional features require comprehensive spatial features such as POI distribution and demographics.

Spatio‑Temporal Representation Generation

The authors propose an Amortization Network that models the approximate posterior of spatio‑temporal representations for target regions and date types, considering both sales and related spatial features.

Temporal patterns, including linear and nonlinear correlations and weekly cycles, are captured using a linear transformation module and a special skip‑LSTM layer, followed by a feature‑fusion module that combines temporal and static features.

Spatio‑Temporal Sales Prediction

A generative model leverages shared statistical structure to learn across tasks. It combines two inputs: (1) a recent‑sales representation extracted via feature extraction and fusion, and (2) a task‑specific spatio‑temporal feature sampled from the amortization network. A fully‑connected network merges them to predict future sales orders.

Spatio‑Temporal Alternating Training

Inspired by multi‑view learning, two views—spatial and temporal—are constructed. Alternating training switches between views, allowing the model to learn complementary information and integrate the shared generative model, improving prediction accuracy.

Experiments

Evaluations on different time intervals (weekday, weekend, shopping festivals) and ablation studies show that the proposed STMP model outperforms baselines, reducing MSE by at least 30 % and achieving superior performance in bursty sales scenarios.

Further analysis of model variants demonstrates the importance of task‑specific spatio‑temporal distributions and modeling multiple periodic patterns.

Visualization on four representative regions across 2019 confirms that STMP accurately captures explosive sales peaks that baseline methods miss.