Regime-Dependent Performance of ARIMA and Modern Forecasting Methods: An Empirical Benchmark on Small-Scale Retail Demand Data

Abstract

Empirical benchmark comparing ARIMA and modern forecasting methods across demand regimes on small-scale retail data, revealing that performance differences are regime-dependent rather than universal.

Methodology

1. 01Segmented demand data by regime: stable, seasonal, volatile, sparse
2. 02Benchmarked ARIMA, ETS, LightGBM, LSTM, and TFT across all regimes
3. 03Applied statistical significance testing to regime-performance interactions

Tools Used