Demand Forecaster
An interactive forecasting model trained on synthetic retail data. Adjust inputs (season, price, promotions, day of week) and see predicted daily unit sales shift in real time. Shows predicted vs actual on historical data, plus 30-day forward projection with confidence intervals.
Model
Gradient Boosting
MAPE
8.2%
Inference
< 20ms
Dependencies
None (local model)
How it works
Model Architecture
Gradient Boosting was chosen over linear regression (can't capture non-linear seasonal patterns) and neural networks (overkill for tabular data, needs GPU). Seven engineered features: price, day of week, month, is_promo, lag_7, rolling_14_mean, and is_holiday (UK bank holidays).
Evaluation
MAPE 8.2% on test set, R² 0.89, MAE 12.4 units/day. Confidence intervals come from quantile regression — separate models trained for 10th and 90th percentiles. TimeSeriesSplit cross-validation ensures no future data leaks into training.
Security
- Model integrity verified via SHA-256 checksum
- Input validation via Pydantic (price >0 and <10k, dates within range)
- Forecast horizon capped at 90 days
- Rate limiting: 20 predictions/min per IP
- No external dependencies at inference time
Want something like this built for your business?
I'll look at your problem, figure out the right approach, and ship working software. No slideshows.
Book a free consultation