They were about to buy a billion-dollar company on a number nobody trusted.
A private equity team had to decide: acquire, or walk. Published commodity forecasts and a leading neural net both missed. Olivet cut the error to about a third, with confidence bands they could take to committee.
A private equity team was looking at a billion-dollar business whose earnings moved with a commodity price. Buy, or walk. The investment committee was not going to sign off on “the industry newsletter said so.” Firm, target, and market stay off this page. The question was blunt anyway: where does this price go, and how sure are you?
They already had what everyone else had. Published sheets. Consensus. Then a leading neural net they had built in-house: stacked layers, hundreds of training rounds, the whole costume. It still missed. Especially when the series changed character, which is the only time a forecast is worth putting a bid on.
Hide the later years. Then look.
We trained on the older history. Inventories. The price itself. Freight. Exchange rates. A few related markets. Seasonality, which everyone waves at and almost nobody fits honestly. Then we locked the model and scored it on years it had not seen. That is the closest thing to sitting in the committee before the print arrives.
Mean absolute error on that held-out period: the net sat near 100 on our index. Olivet sat at 32. Same target. Same years. Not a comparison on the training set.
HELD-OUT ERROR (LOWER IS BETTER)
Almost all of it was two things
Seasonality first. Freight second. Together they were about 70% of the model. Lagged and related prices helped. Exchange rates and a grab-bag of public company prices barely moved the needle, which is useful to know because those are the series people love to add when they are nervous about a deal.
A partner can underwrite two families of inputs. They cannot underwrite hundreds of opaque weights. That is the whole point of measuring this instead of staring at a training chart.
WHAT CARRIED THE FORECAST
The drop
In the test window the price broke hard. The model went to high confidence and stayed inside its bands. That is the rare trick, and it is the one that matters in diligence. Most systems get loud when they are lost. This one stayed sure, and the actual price cooperated: 89% of actuals landed inside tight intervals over the whole held-out period.
Pattern strength came in at 3.57 out of 5. Good enough to run. More recent data would raise it. We said that out loud. A deal model that cannot say “this is only a 3.6” should not be in the memo.