How confident should we be in the forecast?
Backtest: better model, not accurate model
Across the two vintages this data supports, the driver-based forecast landed closer than a naive extrapolation on all six metric-year pairs. That is a statement about a weak baseline, not about accuracy: both methods undershot in both years, and the FY2024 operating-profit forecast was wrong by 63%. Beating a naive baseline is not the same claim as being reliable.
| Naive | Driver-based | Actual | |
|---|---|---|---|
| Revenue | €26.18bn | €25.59bn | €24.81bn |
| Operating profit | €1.60bn | €1.75bn | €2.06bn |
| Free cash flow | €1.27bn | €1.07bn | €1.11bn |
Why it missed
The scorecard says by how much. This says which assumption was wrong — each driver walked from what was forecast to what was reported, one at a time, with whatever the drivers do not explain left visible as a residual rather than absorbed into the last step. Sequential bridge: drivers are substituted in a fixed order, so interaction effects are attributed to whichever driver moves later.
The same assumption dominates both years, in opposite directions. Working capital was forecast at 23.5% of sales for FY2024 and came in at 19.7% — a release worth +€900m of cash nobody planned for. A year later it was forecast at 21.5% and came in at 23.0%, a build worth −€372m. The driver this model ranks first for management attention is the one that has actually driven the forecast error, twice.
FY2023 → FY2024
| Driver | Assumed | Actual | Impact |
|---|---|---|---|
| Revenue growth net sales over the prior year, as reported | 5.0% | 10.5% | −€201m |
| EBITDA margin EBITDA over net sales, as reported | 7.3% | 10.4% | +€481m |
| Effective tax rate effective tax rate, as reported | 34.5% | 26.5% | +€101m |
| Working capital average operating working capital as a percentage of net sales, as disclosed | 23.5% | 19.7% | +€900m |
| Capex capital expenditure, as reported | €600m | €540m | +€60m |
| Residual not attributable to any driver | — | — | −€20m |
Not attributable to a driver: the model scales D&A with revenue, product-division revenue does not sum exactly to reported group net sales, and free cash flow is a derived construction rather than a disclosed line item.
Sequential bridge: drivers are substituted in a fixed order, so interaction effects are attributed to whichever driver moves later.
FY2024 → FY2025
Driver errors largely offset: €808m of gross movement nets to €+37m. The forecast landed close on this metric despite every assumption behind it being wrong, so the small variance is not evidence the assumptions were sound.
| Driver | Assumed | Actual | Impact |
|---|---|---|---|
| Revenue growth net sales over the prior year, as reported | 8.0% | 4.8% | +€90m |
| EBITDA margin EBITDA over net sales, as reported | 11.6% | 12.6% | +€181m |
| Effective tax rate effective tax rate, as reported | 26.5% | 24.3% | +€43m |
| Working capital average operating working capital as a percentage of net sales, as disclosed | 21.5% | 23.0% | −€372m |
| Capex capital expenditure, as reported | €600m | €477m | +€123m |
| Residual not attributable to any driver | — | — | −€27m |
Driver errors largely offset: €808m of gross movement nets to €+37m. The forecast landed close on this metric despite every assumption behind it being wrong, so the small variance is not evidence the assumptions were sound.
Not attributable to a driver: the model scales D&A with revenue, product-division revenue does not sum exactly to reported group net sales, and free cash flow is a derived construction rather than a disclosed line item.
Sequential bridge: drivers are substituted in a fixed order, so interaction effects are attributed to whichever driver moves later.
Does the error repeat?
Two vintages, each built from the initial guidance in the prior year’s report — never a figure revised part-way through the year it describes. One point shows the size of a miss; two show whether it recurs. Both years undershoot operating profit, because adidas guided conservatively in both and beat its own guidance in both. That is a property of the input, not of the arithmetic.
| Error vs. actual | FY2023 → FY2024guidance published February 2024 | FY2024 → FY2025guidance published March 2025 | ||
|---|---|---|---|---|
| Driver-based | Naive | Driver-based | Naive | |
| Revenue | −5.0% | −13.9% | +3.1% | +5.5% |
| Operating profit | −62.6% | −84.8% | −14.9% | −22.3% |
| Free cash flow | −54.7% | −72.2% | −3.4% | +14.8% |
Winning every comparison is a statement about the baseline, not about accuracy. A −62.6% miss beats a −84.8% miss and both are badly wrong. What these two years support is that the error repeats: guidance-anchored forecasts inherit whatever conservatism the guidance carried.
Monte Carlo range
Ranges are adidas's own disclosed FY2025 guidance bands, not historical volatility -- three fiscal years is too few to estimate volatility honestly. Capex has no disclosed range; a +/-5% band was assumed around the guided point figure.
Which assumption explains the most variance?
The simulation ranks assumptions by how much of the free-cash-flow spread each one explains. That is a statistical reading, and it is not the same question as where management should spend its attention — an assumption can dominate the variance and still be one nobody can move. Priorities answers that second question.
- Working capital, % of sales-0.92
- Operating profit within guidance0.27
- Capital expenditure-0.22
- Revenue growth-0.19
Data lineage — an example
What this is not
- A multi-company benchmark. One business, modelled in depth over a single planning horizon — not a peer comparison.
- A price/volume analysis. adidas does not disclose that split, so product-division growth is the driver used here. The channel split (wholesale and direct-to-consumer) is extracted and available as an alternative segmentation, but the model runs one at a time.
- A track record. Two backtest points. Enough to show an error repeating and to expose a broken assumption; not enough to show the approach generalizes, and both forecasts missed by a lot.
- A trading or investment signal. A methodology exercise on public financial disclosures.
- Advice. Nothing here is a recommendation. The management questions are prompts for a discussion, not conclusions from one.