Does machine learning forecast lumber prices better than the market?

The four recent studies we read don't say, at least in their abstracts, because none compares its models with the market. A 2026 study forecast daily lumber prices with Gaussian process regression and reported a relative root mean square error of 3.32% over two years out of sample. A 2023 study of roundwood prices found a neural network "far superior" to classical models on price changes and levels, though error measures varied by species. A 2022 study found deep learning beat other machine learning on lumber futures. None of the four abstracts reports how the models did against the forecast that tomorrow's price equals today's, or against the futures price.

Source: Jin and Xu, "Machine learning lumber price forecasts", International Journal of Financial Engineering, published 5 January 2026. Read from its abstract; the full text is behind a paywall, and it may contain benchmarks the abstract omits.

What the usual answer says

The top results are the lumber futures nowcasting study, a 2021 neural network study and the 2026 paper. The usual answer is that machine learning forecasts lumber prices with low error and often beats traditional methods. Individual papers do report that. Beating a traditional model is a different claim from beating the market.

What the studies found

Jin and Xu used "the lumber price that is released every day", from "January 2, 2014 to April 30, 2024". Their abstract doesn't say whether that is a futures or a cash price series. They tested from "April 15, 2022 to April 30, 2024". Their result: "the relative root mean square error for the price of lumber was 3.3216%". They call the estimates "reasonably accurate".

The 2023 roundwood study by Kożuch, Cywicka and Adamowicz compared neural networks with classical models on quarterly prices from 2005 to 2021. Its headline is clear. "MLP was found to be far superior to other models in terms of forecasting price changes and levels." Classical models "had a tendency to smooth price trends and produce forecasts biased toward average values". But "Ex-post error-based measures of prediction accuracy revealed a complex picture". Exponential smoothing did best for one species and other classical models for another.

He, Li, Via and Zhang's 2022 study used internet search data to nowcast lumber futures. It found that "deep learning models can better capture trends and provide more accurate predictions than machine learning models". A 2021 study used a recurrent network on lumber prices through two recessions. It reported that the network "was able to capture the trend for both recession periods with a remarkably low degree of error".

Reading it through a data lens

In our view, a price model has two easy rivals. One is the forecast that tomorrow's price equals today's. The other, for any horizon beyond a day, is the futures price, if the series being forecast isn't itself a futures price. A model earns trust by beating both over the same test period. All four studies compare models with models, which says which model is best among those tried, not whether any of them adds to what the price already shows.

The test window matters too. A low error over one stretch of prices may not survive a sudden spike. The 2021 study's choice of two recessions is a useful exception. So is the 2023 study's finding that the best method changed from one species to the next. A buyer comparing models across products should expect the same, and test each product's series separately rather than trusting one headline figure.

A worked check

Say a dealer wants to test a vendor's lumber price model over the last two years. It computes three errors for next-week forecasts: the model's, a no-change forecast's, and the nearest futures price's. If the model beats both, it has earned a trial. If it beats the no-change forecast but not futures, the market already knew what the model knows. If it beats neither, the low headline error came from the series, not the model.

What we cannot take from it

We read abstracts, so the papers may report benchmarks we can't see. The 2023 study covers roundwood, not lumber. None of the studies measures whether traders or mills using the models did better.

Quarri for finance and strategy teams is built for the people who close the month, explain the margin and answer the board.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.