Blog · Procurement and supply chain · Logistics
Compare logistics volume and transit forecasts with actual outcomes before using them in planning. This worked example measures error, bias, and performance against a simple baseline, with checks on forecast timing and shipment coverage.
Predictive analytics for logistics estimates future shipment volumes and transit outcomes, then tests those estimates against what happened. The useful question is whether the forecast improves a planning decision compared with a simpler alternative. Preserve the forecast issued before the outcome and compare both approaches on the same shipments or periods.
A volume forecast might support weekly labor scheduling or purchased-capacity planning. A transit estimate might help an account team assess whether an existing delivery promise is realistic. These decisions need different units, deadlines, and tolerances.
Write down the quantity being predicted: tendered loads, completed shipments, elapsed transit hours, or arrival date. Specify when the estimate must be available. A forecast updated after the scheduling deadline cannot demonstrate that the original staffing decision would have improved.
The broader predictive analytics guide explains the approach across business functions. Here the job is narrower: evaluate the predictions already present in a logistics export.
For volume, retain one row per lane, service class, target week, forecast issue date, and forecast method. For transit, use shipment ID plus the issue timestamp so later arrival updates do not overwrite the original estimate. Add actual quantities only when the corresponding period or shipment has reached the agreed evaluation cutoff.
Keep the information available at the time separate from facts learned afterward. The authors of Forecasting: Principles and Practice distinguish test-set forecast errors from training residuals: fitting historical observations well does not establish performance on unseen outcomes.
Join each forecast to its matching actual once. Report unmatched forecasts, duplicate shipment IDs, and missing delivery confirmations before interpreting any score. An undelivered shipment is unresolved, not a completed delivery with no error.
An illustrative US freight broker tests forecasts for a Chicago outbound operation. Each forecast was issued before its target week. The baseline uses the preceding completed week's actual shipments, shown explicitly in the table.
Here, signed error means forecast minus actual. A negative value means underprediction. All methods use the same completed weeks.
| Target week beginning | Forecast shipments | Actual shipments | Signed error | Absolute error | Baseline shipments | Baseline absolute error |
|---|---|---|---|---|---|---|
| August 3, 2026 | 100 | 110 | -10 | 10 | 95 | 15 |
| August 10, 2026 | 120 | 100 | 20 | 20 | 110 | 10 |
| August 17, 2026 | 110 | 120 | -10 | 10 | 100 | 20 |
| August 24, 2026 | 130 | 150 | -20 | 20 | 120 | 30 |
| Total | 460 | 480 | -20 | 60 | 425 | 75 |
Mean absolute error (MAE) = Sum of absolute errors / Number of evaluated periods
Weighted absolute percentage error (WAPE) = Sum of absolute errors / Sum of actual volume
Volume bias percentage = Sum of signed errors / Sum of actual volume
MAE is 60 / 4 = 15 shipments per week. WAPE is 60 / 480 = 12.5%. Bias is -20 / 480 = -4.2%, rounded to one decimal place. The forecast underpredicts total volume, even though it overpredicts the second week.
The baseline MAE is 75 / 4 = 18.75 shipments. The reduction in MAE is (18.75 - 15) / 18.75 = 20.0%. That is an improvement on these rows; the short example does not establish how the method will perform across a full season.
The arithmetic checks are separate: 460 - 480 = -20 proves signed error, while 10 + 20 + 10 + 20 = 60 proves absolute error. Using only the net difference hides offsetting misses.
Volume accuracy does not establish delivery-time accuracy. Keep elapsed hours on a consistent clock, including an explicit treatment of weekends and time zones. Retain the customer's original agreed deadline alongside the transit prediction.
| Shipment | Predicted transit hours | Actual transit hours | Agreed deadline after pickup, hours | Signed error, hours | Absolute error, hours |
|---|---|---|---|---|---|
| T1 | 24 | 30 | 24 | -6 | 6 |
| T2 | 36 | 30 | 36 | 6 | 6 |
| T3 | 24 | 24 | 30 | 0 | 0 |
| T4 | 48 | 60 | 60 | -12 | 12 |
| Total | 132 | 144 | 150 | -12 | 24 |
Transit MAE is 24 / 4 = 6 hours. Mean signed error is -12 / 4 = -3 hours, so the predictions are early on average. Under a declared evaluation tolerance of plus or minus 3 hours, only T3 qualifies: 1 / 4 = 25.0%.
Yet T2, T3, and T4 meet their agreed deadlines: 3 / 4 = 75.0% on-time delivery. T4 is late relative to its prediction but meets its customer commitment. These scores answer different questions.
The Federal Highway Administration explains travel-time reliability as consistency or dependability, which averages alone can conceal. For shipment analysis, inspect the individual misses and their lane and service context alongside the mean. Use data analytics in logistics for the related comparison of observed service and shipment contribution.
Repeat the test across successive forecast issue dates, always evaluating the horizon the business actually uses. Time-series cross-validation evaluates forecasts with training observations that precede each test observation; it avoids letting future outcomes inform an earlier prediction.
The planning manager owns the required horizon and the consequences of underprediction. The transportation manager owns transit definitions and delivery exceptions. Finance should assess whether a measured improvement would justify a change in staffing or capacity commitments, rather than assuming a lower error score automatically creates savings.
For the distinction between describing completed work and estimating future outcomes, see descriptive versus predictive analytics.
Bring your forecast register and matching shipment actuals as a file or Google Sheet; Covirage maps the columns, and its tools compute and check the comparisons from those rows. The external AI model explains the results with every figure traced to its rows. Create a free Covirage account and bring a sample of your own export.
See Covirage for shipping and logistics for the industry context.
Keep the forecast issue date, target period or shipment ID, forecast value, actual value, lane, and service class. Record the unit and forecast horizon too. Retain cancellations and missing outcomes as explicit statuses so they do not silently change the evaluation population.
There is no universal score. Compare with a simple baseline on the same lanes, dates, and planning horizon, then assess the operational cost of the remaining errors. A useful improvement must hold across representative periods and support the decision the planner needs to make.
Yes. Prediction accuracy compares the predicted arrival with the actual arrival. On-time delivery compares the actual arrival with the customer's agreed deadline. A prediction can accurately anticipate a missed commitment, so report forecast performance and customer service separately.