Client€5B Nordic Energy Provider
SprintAn 8 week sprint
Fieldwork4 commodities · 5 years of quarterly data · 5 regression models
Strategy and Commercial Excellence
The acquisition came with a trading book
A utility had acquired a business whose economics it did not share. Most of the acquired revenue came from commodity trading, which moves with oil, gas, coal and carbon prices rather than with anything the parent recognised. The tax position depended on income nobody could forecast. So the sprint built something that could forecast it, and was honest about how far it could be trusted.
0.74
Best model fit, against a stated threshold of 0.70
2 of 4
Models that cleared that threshold
20
Quarters of price and income data behind the regressions
±30%
Stated accuracy of the finished model
A note on sourcing. Every figure here comes from a SprintlyWorks client engagement. Clients are described, never named. Where a figure is identified, modelled or indicative rather than banked, the line says so.
01
A parent that cannot forecast its own consolidated income
The acquired business sells into commodity markets. More than 85 per cent of its revenue comes from energy trading, on its own published figure. That revenue moves with prices the parent had never had to model, and the tax position for the combined group moves with it.
The question put to the sprint had three parts. What does the acquired trading income do to the group tax position, how does the trading position change with commodity prices, and what is a rational way to estimate profit for tax purposes and under the accounting standard.
Five years of quarterly price and income data were assembled across four commodities: oil, gas, coal and carbon allowances. Electricity was tracked as a fifth series.
02
What the data actually supported
Four regressions were built, on revenue, other operating income, cost of materials and other operating expenses. The team set its own quality threshold at 0.70 and then reported against it rather than around it. Revenue reached 0.74 and cost of materials 0.70. Other operating income came in at 0.64 and other operating expenses at 0.59.
Two of the four models cleared the bar. That is written down here for the same reason it was written down there: a forecasting chain is only as good as its weakest link, and the weakest links in this one are on the cost side.
Electricity prices were tested and then removed. Adding them produced sensible looking coefficients and made the model worse at predicting, so they came out. Deciding what not to include is most of the work in an exercise like this.
Model quality
Four models, two of them good enough
Ordinary least squares fits on 20 quarters of data. The team own threshold for an acceptable fit was 0.70, and it is reported against rather than hidden.
| What it predicts | R squared | Clears the 0.70 threshold |
|---|---|---|
| Revenue | 0.74 | Yes |
| Cost of materials | 0.70 | Yes |
| Other operating income | 0.64 | No |
| Other operating expenses | 0.59 | No |
How to read this → The two that fail are both on the cost side, which is why the deliverable ended up being a range rather than a number.
SprintlyWorks analysis
03
Oil moves the wrong way, which is the interesting part
Coal, gas and carbon prices all move with the trading revenue, as anyone would expect. Oil moves against it. When oil prices rise, revenue falls.
Gas carries by far the largest coefficient of the four, so gas is the price to watch if you want an early read on where the income is heading.
Several income and cost lines showed no relationship to commodity prices at all and varied by less than one per cent quarter to quarter. Those were modelled as flat percentages of revenue rather than regressed, which is the right answer and a less impressive looking one. Two quarters of depreciation and impairment were excluded as one off spikes, and the exclusion is stated rather than smoothed away.
04
The output is a range, and it is labelled a range
The model produces three scenarios, best, average and worst, built on the statistical rule that almost all observations sit within three standard deviations of the mean. It reports expected income tax, adjusted operating profit and adjusted net income under each.
The specific euro outputs are not reproduced here. They are model predictions for illustrative input prices on an identifiable group, and publishing them would attribute a forecast tax position to a company that never adopted it.
What can be published is the honesty attached to them. The team stated the model accuracy as plus or minus thirty per cent, and flagged that a fully correct adjusted operating profit would require contract volume detail for derivative fair values that was not available. A model that states its own error band is more useful than one that does not, and considerably rarer.
05
What the client was left holding
A working Excel model that takes four inputs, three months of commodity prices, the effective tax rate and the exchange rate, and returns the forecast chain from revenue through to expected income tax. A user manual. And a five page appendix showing how to rerun the regressions when new quarters arrive, so the model can be maintained without us.
That last part is the deliverable that matters. A forecast handed over is worth one quarter. A method handed over with the instructions for rebuilding it is worth as long as the acquisition lasts.
There is no measured result in this work. No tax was saved, no position was filed and no forecast was proven right or wrong within the engagement. What was produced was the ability to put a number and an error band on a question that previously had neither.
Download the full case study
Have a similar requirement?
Contact us today to learn more about on-demand workforce and accelerate development on your most pivotal projects!
Quick Reads for Big Impact
Featured Case Studies
Quick Reads for Big Impact
Accelerating Success for Enterprises in 20+ Geographies
Launch Your Sprint with
Define your project, connect with top-tier consultants, and start making progress fast.