Demand forecasting that cut a logistics operator's idle fleet by a quarter
Fleet allocation ran on a planner's intuition and last week's numbers. A forecasting model made the same decision with two weeks of foresight.
-24%
Idle fleet capacity
Across the two quarters after rollout.
8.6%
Forecast accuracy (MAPE)
At fourteen-day horizon, against a 21% seasonal naive baseline.
-31%
Turned-away demand
Capacity repositioned ahead of predicted regional peaks.
2 days to 3 hours
Planning cycle time
Weekly allocation planning effort.
What was actually going wrong
- Industry
- Logistics
- Services
- AI/ML
- Technologies
- PythonXGBoostAirflowdbtMLflowSnowflake
- 1
Vehicles were allocated to routes weekly using spreadsheets and experience.
- 2
Idle capacity in one region regularly coincided with turned-away demand in another.
- 3
Booking, telemetry and billing data lived in three systems that had never been joined.
- 4
A previous forecasting attempt had been abandoned because nobody trusted its output.
What we did, and why in that order
Sequence mattered here as much as the individual decisions. Each step below made the next one safe.
- 01
Build the data foundation first
Airflow and dbt pipelines unified booking, telemetry and billing into a modelled warehouse with tested data quality expectations, which is where the previous attempt had failed.
- 02
Start with a baseline anyone can argue with
A seasonal naive forecast set the bar. Every subsequent model had to beat it on business cost, not on error metrics alone.
- 03
Forecast at the level decisions are made
Region and vehicle-class granularity over a fourteen-day horizon, matching how planners actually allocate rather than how the data was shaped.
- 04
Put the forecast where the planner already works
Output surfaced inside the existing allocation tool with a plain-language explanation of the main drivers behind each number.
- 05
Monitor for drift
Weekly backtesting against actuals with an alert when accuracy degrades, plus scheduled retraining.
What changed, measured
Each figure below was measured after the engagement, not projected during it.
24%
less idle fleet capacity
Idle fleet capacity
Across the two quarters after rollout.
-24%
Forecast accuracy (MAPE)
At fourteen-day horizon, against a 21% seasonal naive baseline.
8.6%
Turned-away demand
Capacity repositioned ahead of predicted regional peaks.
-31%
Planning cycle time
Weekly allocation planning effort.
2 days to 3 hours
The capabilities behind this engagement
Facing something similar?
We will tell you what we would repeat from this engagement, what we would change, and what it is likely to take.