Australian researchers predict risky driving hotspots with ARIMA beating deep learning

Road safety has long been reactive—waiting for crashes to happen before identifying dangerous roads. Researchers led by Adriana-Simona Mihăiţă at University of Technology Sydney flipped that script using connected vehicle telemetry from thousands of cars in Greater Sydney. By detecting near-miss events via g-force thresholds—hard braking above 0.6g, harsh cornering above 0.47g, and harsh acceleration above 0.5g—they built spatio-temporal heatmaps of risky driving across Local Government Areas (LGAs). The study, submitted to the ATRF 2026 conference, pinpoints Sydney CBD, Parramatta, and Bankstown as persistent high-risk zones requiring targeted policy action.

The team benchmarked eight predictive models across three families: ensemble learners (Random Forests, XGBoost, LightGBM), deep learning (LSTM, N-BEATS), and classical time-series (ARIMA, Exponential Smoothing, Prophet). Surprisingly, the simplest model won. ARIMA achieved the lowest mean absolute error at 162.21, narrowly edging out LSTM (163.92) and handily beating N-BEATS (180.75). This suggests that when training data is limited, parsimonious statistical models are just as effective as computationally expensive deep networks. The findings open the door for IoT-based proactive safety systems that alert authorities before accidents occur, potentially reshaping road safety investment and urban planning across Australia.