DETECTS DEMAND SHIFTS 3 WEEKS EARLY FORECASTING AI PYTHON

Demand Sensing Dashboard

Python + ML pipeline ingesting POS data, social sentiment, weather patterns, and promotional calendars to generate 12-week demand sensing forecasts with confidence bands, anomaly detection, and signal attribution — detecting shifts 3 weeks before traditional models.

Project Overview

⚑ Problem

Lagging Statistical Forecasts

Traditional forecasting relies on historical averages and moving windows — it can only react after demand has shifted. Promotions, weather events, social virality, and competitor actions create demand spikes that arrive 2–4 weeks before they appear in sales data, leaving planners perpetually behind.

⚙ Solution

AI Demand Sensing Pipeline

A Python-based sensing engine integrates POS velocity, Google Trends signals, weather API inputs, promotional calendars, and social sentiment scoring. An ensemble model (XGBoost + Prophet) blends statistical and signal-based forecasts, generating 12-week ahead predictions with confidence intervals and anomaly flags.

✓ Findings

Error Reduced 38% vs Statistical

AI sensing cut average MAPE from 18.4% to 11.4% — a 38% improvement — across 8 SKUs over 12 weeks. Promotional lifts were detected 3 weeks early in 7 of 8 cases. 4 anomaly events were flagged before impacting service levels, avoiding an estimated €340K in emergency procurement costs.

Live Demand Signal Feed

Active Demand Signals — Ingestion Pipeline

LIVE SIGNALS

⚡ SKU Drill-Down — AI vs Statistical Forecast

Dataset Explorer

Forecast Performance Analysis

Actual vs Statistical vs AI Sensing (Filtered SKUs)

MAPE: AI Sensing vs Statistical — by Category

Signal Strength by Week

Forecast Error Distribution

Anomaly Timeline

Confidence % by Week & SKU

Weekly Forecast Dataset

WEEK	SKU	CATEGORY	ACTUAL SALES	STAT. FORECAST	AI SENSING	SIGNAL TYPE	SIGNAL STRENGTH	CONFIDENCE %	AI ERROR %	STAT ERROR %	ANOMALY

8-Week Forward Demand Sensing Forecast

Portfolio Demand — Historical + 8-Week Forward with Confidence Bands

Company Advice

💡 Recommendations for Demand Planning Teams

Demand sensing reduces forecast error by 30–50% at the weekly level versus statistical baselines — the greatest gains are in promotional and weather-sensitive categories where exogenous signals dominate demand.
Don't replace your statistical forecast — layer sensing on top of it. The ensemble approach (signal-weighted blending) consistently outperforms either model alone, especially at short (1–4 week) horizons.
Prioritise signal feeds in order of value: promotional calendars first (highest, most controllable), then weather API, then social sentiment. Competitor signals are high value but harder to source reliably at scale.
Anomaly detection is your early warning system — configure alerts for any week where AI sensing diverges from statistical forecast by more than 15%. Investigate before the week arrives, not after stockout.
Confidence bands narrow as signals strengthen. Share the confidence % with inventory planners — a low-confidence high-uplift week demands safety stock buffer; a high-confidence forecast warrants JIT replenishment.
Refresh the model on a rolling 13-week window. Demand patterns drift, and a sensing model trained on stale data can perform worse than a simple moving average within 6–8 weeks of a regime change.

Key Takeaway: The real value of demand sensing is not better numbers on a spreadsheet — it's the 3-week lead time it gives procurement and production to act before a demand wave hits. Speed of signal is worth more than model accuracy.