ONE HOUR AHEAD / 24 HOURS BEHIND

Forecast with the past.

A small residual network for the next hour of recorded bike rentals. Give it 24 consecutive hours; inspect its estimate beside two simple baselines.

用连续24小时的历史计数预测下一小时。模型来自2011–2012年华盛顿数据，保留时间缺口和完整测试分母；不代表当前需求、可用车辆或预测区间。

Historical experiment: timestamp labels have no timezone. The app uses nominal calendar hours without reconstructing DST or UTC. Inputs stay in this tab unless you download a report.

4,258Eligible test hours / 4,376

33.85Neural mean absolute error

86.91Last-hour baseline MAE

80.94Lag-24 baseline MAE

Try a historical context

Exactly 24 records, oldest first. Use YYYY-MM-DDTHH:00:00 and integer counts from 0 to 100,000. No gap filling or timezone conversion.

Hourly history JSON

Loading fixed trained weights…

SELECTED ON VALIDATION ONLY

One small trained model

35 inputs → 32 tanh units → 1 residual output. 1,185 parameters, 40 actual training epochs; validation selected epoch 20.

The model uses earlier rental counts and the target hour, weekday and month. Target-time weather, casual rentals and registered rentals are excluded.

Test evaluation is rolling one hour ahead: later forecasts may use actual earlier test counts. It is not a six-month forecast from one fixed starting point.

Missing hours and errors

The original hourly CSV contains 17,379 rows and 165 absent hours. A target needs all 24 earlier nominal hours. This removes 1,366 training, 238 validation and 118 test targets. Omitted windows may represent harder periods.

The largest test error is about 395 rentals. Every eligible test prediction and all omitted-row reasons are included in the release files.

Read the protocol and full results

Forty epochs, complete test comparison

All 40 training and validation MAE values, six test-month MAE comparisons for the neural model and both baselines, and all eligible targets by split

These are descriptive results from one historical split and one neural fit. Overlapping windows and the changing seasons do not support an independent-trial or significance claim.

Open full-resolution figure

Forecast with the past.

Try a historical context

Next-hour estimate

Forty epochs, complete test comparison