A retailer who uses sales forecasts to decide on price changes might be misled. Forecasting models predict the sales that typically happen in the past, but a price change asks what sales would be if only price changed and everything else stayed the same. These are different questions because past prices are often set based on demand that the data does not record.
The researchers tested several large forecasting models that are pre-trained on broad data, including ones specifically trained to estimate cause and effect. They found that a model that predicts sales accurately can still get the effect of price completely wrong. Adding more data improved the forecasts but did not fix the price effect.
The main finding is that forecasting accuracy alone is not enough to trust a model for pricing decisions. The error in the predicted effect of price was large even for models built for causal inference. This matters because businesses relying on forecasts for price optimization risk making bad choices.