Model accuracy · Updated monthly

Monthly validation of truck parking predictions

Every month, we compare our predictions with what drivers report on the ground. The results are published here and used to improve the model.

— Latest average error, in points. Lower is better —
— Predictions within 20 points Close to one step of the report scale
— Predictions within 10 points Stricter than the 25-point report scale

Reading the numbers

How to read the numbers

Each driver report is compared with the model’s prediction for the same lot and time.

Average error

Average gap between the prediction and the driver report, in percentage points. Lower is better.

Within 20 points

Share of predictions within 20 points of the driver report. The main figure to watch.

Within 10 points

Share of predictions within 10 points of the driver report. A stricter test.

Example: a driver reports 50%. A prediction of 40% to 60% is within 10 points. A prediction of 30% to 70% is within 20 points.

Results

Monthly results

Month Average error Within 10 points Within 20 points
Jan 202618.0%42.1%59.6%
Feb 202616.7%51.9%72.2%
Mar 202622.7%36.7%53.1%
Apr 202623.8%20.8%56.6%
May 202626.1%22.5%47.5%
Jun 202623.7%16.7%50.0%

Interpretation note: driver reports are not a random sample of all parking conditions. Drivers may be more likely to submit a report when the displayed prediction does not match what they see.

Scale note: drivers report in 25-point steps (25%, 50%, 75%, 100%) while the model predicts any value. A prediction of 63% against a report of 75% counts as a 12-point error, so part of each month’s error comes from the report scale, not the model.

Coverage

Prediction coverage and evidence

Every location has a weekly prediction pattern, but the amount of direct evidence differs by site.

11,791

Total locations. Every location has a complete weekly pattern covering 672 fifteen-minute periods.

6,469 · 54.9%

Site-differentiated. These patterns include direct observations from the individual site.

5,322 · 45.1%

Cluster-inferred. These patterns rely more heavily on operationally similar facilities.

How the prediction baseline works

Locations are grouped using characteristics such as capacity, facility type, fees, amenities, and nearby truck traffic.

Each group supplies a typical weekly availability pattern. Direct driver observations are then used to check or adjust individual locations where evidence is available.

How monthly validation works

Each driver report is matched to the model prediction for the same parking location and the same fifteen-minute period.

  • The model predicts a continuous availability value from 0% to 99%.
  • Drivers report availability in 25-point steps (25%, 50%, 75%, 100%). Reports above 100% are treated as 100%.
  • Average error is the mean absolute gap between prediction and report, in percentage points.
  • Within 10 points shows tighter agreement.
  • Within 20 points shows broader agreement.

Summary

The bottom line

The model provides national truck parking availability patterns without requiring sensors at every facility. Monthly validation makes performance visible, while facility-level confidence information helps planners decide where the data is appropriate for screening or deeper investment analysis.

The trend across months is a more reliable signal of model performance than any single month’s figure.

Authored, reviewed, and approved by Troy Choi, Ph.D., P.E. Transportation Systems Optimization and Engineering Research.

Need confidence-graded parking analytics?

Structured availability patterns, evidence-density information, and corridor-level analytics for freight plans and truck parking studies.