Demand Forecasting Pipeline
LSTM- and SARIMA-based demand forecasting pipeline for industrial R&D, with rigorous evaluation informing operational feasibility decisions.
Overview
Built and evaluated a forecasting pipeline for demand prediction in an industrial R&D context at HELLA InnovationLAB. The project assessed the feasibility of short-term planning based on historical data.
Approach
Two modelling families were compared:
- LSTM (Long Short-Term Memory): Trained on multi-variate time-series data with rolling evaluation windows.
- (S)ARIMA: Classical statistical baseline for comparison and interpretability.
Results
Historical evaluation showed that a three-month forecasting window produced approximately ±5% deviation. This assessment directly informed an operational feasibility decision — the forecasting potential was documented and communicated to stakeholders as a basis for planning.
Limitations
Short-term forecasts showed high variance. The project concluded that operational reliance on forecasts at this horizon required further data collection and model iteration before deployment.
Key takeaway
Rigorous evaluation methodology — not just model accuracy — is what makes a forecasting project useful. Communicating limitations clearly is part of the deliverable.