Model lab

Baseline forecast evaluation

Honest short-horizon baselines establish a comparison bar before Neural ODE modeling. Forecasts use rolling-origin backtests on weekly weighted activity index for Illinois and Cook County. Intervals reflect residual spread from past forecast errors, not clinical uncertainty.

Model runs

ensemble_v1

production

ensemble · v1.0.0

Evaluations scored: 476

Updated Jun 1

moving_average_v1

production

moving_average · v1.0.0

Evaluations scored: 476

Updated Jun 1

neural_ode_17031

production

neural_ode · v1.0.0

Evaluations scored: 237

Updated Jun 1

neural_ode_17031

candidate

neural_ode · v1.2.1

Evaluations scored: 48

Updated May 20

neural_ode_17031

candidate

neural_ode · v1.6.0

Evaluations scored: 48

Updated May 21

neural_ode_17031

candidate

neural_ode · v1.7.3-hardbound-linear

Evaluations scored: 230

Updated Jun 2

neural_ode_17031

candidate

neural_ode · v1.7.2-hardbound

Evaluations scored: 230

Updated Jun 1

neural_ode_17031

candidate

neural_ode · v1.8.0-reliability

Evaluations scored: 226

Updated Jun 1

neural_ode_17031

candidate

neural_ode · v1.1.0

Evaluations scored: 48

Updated May 20

neural_ode_17031

candidate

neural_ode · v1.7.1

Evaluations scored: 230

Updated Jun 1

neural_ode_17031

candidate

neural_ode · v1.3.0

Evaluations scored: 48

Updated May 21

neural_ode_17031

candidate

neural_ode · v1.4.0

Evaluations scored: 48

Updated May 21

neural_ode_17031

candidate

neural_ode · v1.2.0

Evaluations scored: 48

Updated May 20

neural_ode_17031

candidate

neural_ode · v1.5.0

Evaluations scored: 48

Updated May 21

neural_ode_IL

candidate

neural_ode · v1.7.3-hardbound-linear

Evaluations scored: 230

Updated Jun 2

neural_ode_IL

candidate

neural_ode · v1.1.0

Evaluations scored: 48

Updated May 20

neural_ode_IL

candidate

neural_ode · v1.2.0

Evaluations scored: 48

Updated May 20

neural_ode_IL

candidate

neural_ode · v1.2.1

Evaluations scored: 48

Updated May 20

neural_ode_IL

candidate

neural_ode · v1.3.0

Evaluations scored: 48

Updated May 21

neural_ode_IL

candidate

neural_ode · v1.4.0

Evaluations scored: 48

Updated May 21

neural_ode_IL

production

neural_ode · v1.0.0

Evaluations scored: 237

Updated Jun 1

neural_ode_IL

candidate

neural_ode · v1.5.0

Evaluations scored: 48

Updated May 21

neural_ode_IL

candidate

neural_ode · v1.6.0

Evaluations scored: 48

Updated May 21

neural_ode_IL

candidate

neural_ode · v1.8.0-reliability

Evaluations scored: 226

Updated Jun 1

neural_ode_IL

candidate

neural_ode · v1.7.1

Evaluations scored: 230

Updated Jun 1

neural_ode_IL

candidate

neural_ode · v1.7.2-hardbound

Evaluations scored: 230

Updated Jun 1

persistence_v1

production

persistence · v1.0.0

Evaluations scored: 476

Updated Jun 1

seasonal_naive_v1

production

seasonal_naive · v1.0.0

Evaluations scored: 476

Updated Jun 1

trend_v1

production

trend · v1.0.0

Evaluations scored: 476

Updated Jun 1

Performance metrics

Baseline performance

Rolling-origin backtest metrics by forecast horizon. Lower MAE/RMSE is better; trend accuracy measures direction (rising/falling/stable).

Model	1-week			2-week			3-week			4-week			N
	MAE	RMSE	Trend	MAE	RMSE	Trend	MAE	RMSE	Trend	MAE	RMSE	Trend
ensemble_v1	0.345	0.429	38.5%	0.424	0.522	30.0%	0.504	0.609	39.8%	0.573	0.685	35.3%	476
moving_average_v1	0.421	0.502	31.1%	0.516	0.617	29.2%	0.603	0.720	31.4%	0.686	0.799	25.9%	476
persistence_v1	0.320	0.406	46.7%	0.428	0.517	30.8%	0.520	0.630	26.3%	0.603	0.725	24.1%	476
seasonal_naive_v1	0.756	0.928	44.3%	0.760	0.934	45.0%	0.762	0.937	49.1%	0.769	0.944	48.3%	476
trend_v1	0.386	0.502	42.6%	0.532	0.666	31.7%	0.657	0.830	35.6%	0.761	0.981	41.4%	476

MAE by horizon

Mean absolute error on held-out weekly activity index. Compare models before trusting more complex approaches.

How to read these results

Persistence assumes next week equals this week — often hard to beat at 1-week horizon.
Moving average smooths recent weeks; useful when signal is noisy.
Trend extrapolates recent direction; can overshoot at longer horizons.
Seasonal naive compares to the same week last year (52-week lag).
Ensemble averages component baselines for dashboard display; Neural ODE (Phase 7) must beat these on held-out weeks to be promoted.