Hybrid SARIMA-recurrent neural network approach for road accident forecasting
TL;DR
A hybrid model integrates SARIMA forecasts with RNN-based residual predictions as a methodological extension whose empirical performance is rigorously evaluated against the baseline, confirming its effectiveness in capturing temporal dependencies in the log-residual space.
Abstract
This study investigates the forecasting of daily road traffic accidents using time series models, through a controlled comparative evaluation of baseline SARIMA and Hybrid SARIMA-LTSM frameworks. We employ a Seasonal Autoregressive Integrated Moving Average (SARIMA) model to capture linear and seasonal patterns in accident data from 2010 to 2023. To address potential non-linearities, we extend the approach by modelling SARIMA residuals using Recurrent Neural Networks (RNNs), including bidirectional RNN (BiRNN) and Long Short-Term Memory (LSTM) variants, to develop a hybrid SARIMA-RNN framework. The hybrid model integrates SARIMA forecasts with RNN-based residual predictions as a methodological extension whose empirical performance is rigorously evaluated against the baseline. Within residual-domain model selection, the unidirectional LSTM achieved the lowest RMSE (0.184) and MAPE (1.0966%) among tested architectures, confirming its effectiveness in capturing temporal dependencies in the log-residual space. A hybrid SARIMA-LSTM model was also evaluated to enhance residual prediction; however, the one-step-ahead rolling-origin strategy with a 7-day sliding window did not improve accuracy relative to the baseline SARIMA, yielding only minor increases in RMSE and MAPE.