Halo Tweak Makes AI Forecasts Up to 16% More Accurate
More accurate price and demand predictions could quietly lower your bills.
When an AI predicts something — tomorrow's electricity price, how many people will show up, how much power a city needs — it usually gives one number and stops there. But it never says how confident it is. A new paper from researcher Adam Cataldo shows that simply asking the AI to also report its own uncertainty makes its main prediction noticeably better. He calls the method Halo.
How it works: Halo doesn't build a new model from scratch. It takes an AI that already makes forecasts and adds a second output channel — basically a second dial that reports how spread out the likely answers are, in the same way a weather app says "70% chance of rain" rather than just "rain." That extra information feeds back into training, and the main prediction sharpens as a side effect. Notably, this contradicts earlier findings outside of time-series forecasting, where adding uncertainty estimates often made predictions worse.
Cataldo tested Halo by bolting it onto three very different existing models: a transformer (the same family of AI behind chatbots), a graph network paired with a variational autoencoder, and a simple convolutional network. He ran them across five electricity price markets in a standard forecasting benchmark. Result: accuracy improved in 28 out of 30 model-market-metric comparisons, cutting average squared error by 2.6% to 16.5% and average absolute error by 1.7% to 11%. Two practical findings stood out. First, whether the uncertainty dial is a small add-on or a full parallel network barely matters — what matters is that the model estimates uncertainty at all. Second, the improvement works with the settings already tuned for the old model, so no expensive retuning is needed.
The catch: this is a research paper, not a product you can buy today. It was tested on electricity price forecasting, so gains elsewhere — healthcare, logistics, retail demand — are promising but unproven. Still, the low cost of adoption makes it the kind of quiet upgrade that tends to spread fast.
- Halo teaches AI to say how confident it is, which surprisingly makes its main prediction more accurate too
- Accuracy improved in 28 of 30 tests, cutting average error by roughly 2% to 16% on electricity price forecasts
- It bolts onto existing AI models with no rebuilding or retuning, so the upgrade is cheap and fast to adopt
Why It Matters
Sharper forecasts for energy, demand and prices could mean lower bills, less waste and fewer costly surprises.