Model performance
See where the model is right, wrong, and improving.
Replay held-out periods, compare Flowcast predictions with observed demand and the one-week seasonal baseline, and inspect error by horizon, borough, time, and taxi zone.
How to read this dashboard
Everything here is held-out test data: hours the model never saw while training or tuning. The Forecast page answers “what does the model expect?”; this page answers “how well has it done?”.
- Actual
- Observed yellow-taxi pickups from the public TLC trip records.
- Predicted
- What Flowcast forecast for the same zone and hour, made from information available at the forecast origin.
- Baseline
- The seasonal-naive forecast: the same zone and hour one week earlier. Flowcast has to beat it.
- WAPE
- Total absolute error ÷ total actual pickups. 15% means misses add up to 15% of real demand.
- MAE / RMSE
- Average miss per zone-hour, in pickups. RMSE weighs large misses more heavily.
- Residual
- Actual minus predicted. Positive = Flowcast underpredicted; negative = overpredicted.
Use the filter bar to switch between the full test period and a replay of its last seven days, pick a horizon, and narrow to a borough or a single zone. Every chart, card and table updates together.
- Full held-out test period
- 1h horizon
- Metric: WAPE
- Held-out test period
Observed TLC pickups in the selected frame/filter.
Flowcast’s predicted pickups for the same frame.
Total absolute error ÷ total actual demand. Lower is better.
Average miss per zone-hour, in pickups. Lower is better.
Like MAE, but large misses count more. Lower is better.
Error reduction vs. same zone and hour one week earlier. Positive is better.
Backtest frame
Observed and predicted demand for one held-out timestamp. Scrub or play through the last 7 days of the test period; click a zone to filter the dashboard to it.
Citywide demand over time
Observed pickups vs. Flowcast predictions and the one-week seasonal baseline (daily totals, 1h horizon).
Error by forecast horizon
Lower values indicate more accurate forecasts. Compare 1h, 6h, and 24h on the same evaluation window (WAPE, All boroughs).
Performance by borough
Forecast error across the five boroughs for the current filters (WAPE, 1h).
Yellow-taxi pickups are concentrated in Manhattan and at the airports; outer-borough zones are sparse, so a few pickups can swing their percentage error.
Residual distribution
Residual = actual minus predicted. Values near zero indicate smaller errors; positive values mean Flowcast underpredicted and negative values mean it overpredicted.
Error by hour of day
WAPE by the target hour’s time of day. Shows when in the day the model and the baseline miss most.
Zone performance
Every NYC taxi zone in the current filters, 1h horizon, full held-out test period. Sort any column; select a zone to focus the dashboard and map on it.
Largest forecast misses
Zone-hours with the largest absolute prediction errors in the selected window.
Big misses cluster around unusual hours (events, weather, holidays) that recent history and the weekly pattern do not anticipate.