Driver-based planning promises a budget that moves with the business: instead of negotiating every cost line, finance links each one to the activity that drives it, so that a change in volume flows through to the costs automatically. The promise depends on one thing that is rarely checked: whether the drivers really drive the costs, and by how much.
The usual approach is to take last year's cost, divide it by last year's volume and call the result the variable cost per unit. That assumes every cost is fully variable, which is almost never true. Linear regression, the oldest tool in the statistician's kit, does the job properly, and it has been doing so in industry for a long time: electricity utilities have forecast demand from temperature and calendar effects with regression models for decades.
Fixed and variable, estimated from the data
Take three years of monthly data for a distribution centre: the number of shipments and the total logistics cost. Plotting one against the other and fitting a straight line by least squares separates the two components that the planning model needs.
| Result | Estimate | What it means for the plan |
|---|---|---|
| Intercept | 452 (standard error 44) | Fixed cost per month: rent, supervision, systems. It does not fall when volume falls. |
| Slope | 17.95 (standard error 0.82) | Variable cost per thousand shipments: picking, packing, temporary labour. |
| R² | 0.93 | Shipments explain 93% of the month-to-month variation in cost; the driver is a good one. |
| Residual standard error | 49 | The typical monthly deviation from the line that is not explained by volume. |
The difference from the naive method is material. At the average volume of 53 thousand shipments, the unit-cost approach would treat the whole cost of about 1.4 million as variable, 26.5 per shipment. The regression says only 18 of it is variable. If volume falls 20%, the naive budget cuts cost by 280 thousand a month; the realistic expectation is a cut of about 190, and the gap is a margin shortfall that would otherwise be discovered as an unexplained variance.
Plan with the interval, not just the line
The regression also says how wrong a forecast will be. For a month with 60 thousand shipments, the line gives a cost of about 1,529 thousand, and the 95% prediction interval is roughly 100 thousand either side. That range is the honest answer to "what will logistics cost next month?" and the right basis for judging variances: an actual inside the interval needs no explanation, one outside it does.
Pitfalls to avoid
- Correlation is not a driver. Choose drivers with an operational reason to move the cost, then let the data measure the effect. A regression on an unrelated series can fit well by coincidence.
- Stay inside the range of the data. The model is reliable between about 34 and 68 thousand shipments. Above that, a second shift or a new warehouse would change the cost structure, and the line knows nothing about it.
- Check the residuals over time. If they trend, something else is moving the cost, such as wage inflation, and needs its own driver.
- Use several drivers when the process has them. Multiple regression can separate the effects of shipments, order lines and distance in the same model.
Bartor S.
Founder
