Demand Forecasting & Model Benchmarking
Time-series forecasting pipeline comparing ARIMA/SARIMA baselines with LSTM models, built to establish whether the added model complexity was justified.
Overview
A planning use case required forecasting from noisy historical time-series data. The question worth answering was not “can a model be built” but “is a model good enough to plan against, and does complexity buy anything”.
It ran in two separate passes, roughly a year apart - long enough that the second was a genuine reassessment rather than a continuation.
Approach
First pass - a statistical baseline
ARIMA/SARIMA models via statsmodels. Classical, interpretable and cheap to defend: the point was to establish what a conventional method achieves before reaching for anything heavier.
Second pass - LSTM models
The problem was revisited with LSTM networks in TensorFlow/Keras, trained on multi-variate time-series data with rolling evaluation windows.
Each horizon was benchmarked against baselines rather than against intuition. Having the earlier statistical baseline on hand is what made that judgement possible at all.
What it built
Experience in comparative forecasting, reliability assessment, and treating deployment as a conclusion to be earned rather than the default outcome of a modelling project. Communicating what a model cannot support is part of the deliverable, not a footnote to it.