Methodology
What the model actually does, and where it stops being trustworthy.
1. Ingestion and provenance
Every source is wrapped in a common adapter interface (fetch, normalise, health-check). Each stored observation keeps its source key and provenance: measured, estimated, forecast or demo. This deployment runs entirely on demo data, and the badge on every number says so.
2. Feature engineering
For each lane the engine computes lags (5/10/15/30/60 min), slopes over the same spans, acceleration and jerk, rolling mean and median, an EMA, volatility, data freshness, calendar flags (weekend, SG and MY holidays, school holidays, long weekends, holiday eves) plus incident severity, approach-road congestion and the cross-checkpoint delta.
3. Forecasting
The MVP model blends a comparable-day historical profile (same weekday and holiday character) with current momentum that decays over the horizon. Prediction intervals widen with horizon length, volatility and staleness, and the model emits probabilities of falling under 30, 45, 60 and 90 minutes.
4. Low-window detection
The forecast is scanned for sustained local minima: a window must stay low for at least 25 minutes. Candidates are scored on depth below the current wait, duration, confidence and how soon they start, then ranked.
5. Leave-now vs wait
The decision engine compares leaving immediately against waiting for the best window, requiring a minimum expected saving before it tells anyone to wait, and adds the Tuas detour penalty when it suggests switching checkpoints. It never recommends a switch that costs more driving than it saves queueing.
6. Accuracy and honesty
Backtests replay the model against stored history and report MAE, median error, RMSE, directional accuracy and — most importantly — whether the recommended window actually contained the real minimum. Confidence shown in the UI is derived from data freshness, sample density and volatility, not decoration.
Limitations
Demo data is synthetic and does not reflect real border conditions. Real-world accuracy depends on feed licensing, refresh rate and coverage. Forecasts degrade sharply during unmodelled events such as system outages, enforcement operations or severe accidents.