AI for retail demand sensing: real-time signal processing

Demand sensing uses real-time data signals (POS data, weather, social trends, events, competitor pricing) to continuously update short-horizon forecasts (1–14 days) throughout the planning cycle — often refreshing every few hours. Traditional demand forecasting uses historical data to generate weekly or monthly forward-looking projections that are updated infrequently. Demand sensing detects and responds to demand shifts as they happen, enabling supply chain adjustments before inventory problems develop. The two approaches are complementary: sensing improves short-horizon accuracy while statistical or ML forecasting provides the long-horizon plan.

The highest-impact signals for retail demand sensing are: real-time POS transaction data (the primary signal — hourly sales velocity vs baseline detects demand shifts immediately); weather forecasts (especially for weather-sensitive categories — the correlation between temperature and ice cream/cold beverage sales can be modelled with high accuracy); local events calendar (sporting events, concerts, school calendars drive category-specific spikes); social media virality signals (TikTok and Instagram can drive 10–100× demand spikes within 24 hours — social monitoring provides 1–3 days of advance warning); and competitor promotional activity (significant competitor promotions cause measurable cross-category demand shifts).

Modern demand sensing systems use ensemble approaches combining multiple model types. Gradient boosting models (LightGBM, XGBoost) handle structured tabular signals (weather, promotions, events, historical sales) effectively. LSTM neural networks and Temporal Fusion Transformers capture complex sequential patterns in time series sales data. NLP models process social media and review text signals. Causal models estimate promotional lift. A meta-learner ensemble combines these forecasts, weighting each model by its recent accuracy on similar SKUs. The Temporal Fusion Transformer has emerged as a particularly strong single-model architecture because it handles all input types and provides interpretable attention weights.

Social media monitoring provides leading indicator signals that precede POS spikes by 1–5 days. When a product goes viral on TikTok or appears in a high-reach influencer post, social monitoring tools detect the spike in mentions and sentiment before consumers have visited stores or made purchases. NLP models classify product mentions, assess sentiment, and estimate virality potential. These signals are fed into the demand sensing model to adjust short-horizon forecasts upward for affected SKUs. This 1–3 day advance warning is enough time for targeted replenishment orders or stock transfers to prevent stockouts during the demand surge.

Demand sensing accuracy is measured using MAPE (Mean Absolute Percentage Error), weighted MAPE (giving higher weight to high-volume SKUs), and bias metrics (systematic over- or under-forecasting). Accuracy is measured at multiple horizons (1-day, 3-day, 7-day, 14-day forward) and at multiple granularities (SKU-store, category-store, total store). Compare accuracy against the baseline statistical forecast to quantify the sensing improvement. Track fill rate (stockout reduction), overstock reduction, and days of supply as business outcome KPIs alongside technical accuracy metrics. Planner override rate is also tracked — high override rates indicate low model trust.

Demand sensing requires: real-time POS data streaming infrastructure (Kafka, Azure Event Hub, or AWS Kinesis for sub-hourly POS feed); a feature store (Feast, Tecton, or Databricks Feature Store) that computes and serves derived features for model inference; external data integrations (weather API, events API, social monitoring tool, competitor price scraping); a model serving infrastructure for low-latency forecast inference (typically FastAPI or Ray Serve); a time series database (InfluxDB, TimescaleDB, or Databricks Delta Lake) for storing forecast history; and integration with the planning system (SAP IBP, Kinaxis, o9) via API for forecast consumption by supply chain planners.

Demand sensing delivers the highest ROI for retailers with: high SKU count and complex assortments (grocery, apparel, electronics) where manual forecast management is impossible; short product shelf life where forecast errors directly cause waste (fresh food, perishables); high demand volatility driven by weather, events, or trends (fashion, seasonal products, beverages); large store networks where demand patterns vary significantly by location; and frequent promotions where promotional lift modelling is critical. Retailers with simple, stable demand patterns and long lead times benefit less, as their existing statistical forecasting is already adequate for their planning horizon.

Demand sensing forecasts integrate with automated replenishment systems via API, replacing or blending with the statistical forecast that drives replenishment calculations. In SAP IBP or similar systems, demand sensing short-horizon forecasts replace the statistical consensus forecast for the 1–14 day horizon, while the statistical plan drives longer-horizon planning. Replenishment rules (reorder points, safety stock levels, order quantities) are automatically adjusted based on the sensing forecast. For retailers with vendor-managed inventory (VMI), demand sensing signals are shared directly with suppliers to trigger upstream replenishment before the retailer's own DCs run short.