A forecasting method scores each clue on its own before combining them, improving accuracy.

Forecasting systems that use large language models often gather evidence to predict events like sports or financial markets. Most systems feed all gathered evidence into an AI at once to get a final guess. The authors call this Monolithic PredictionMonolithic PredictionA method where all evidence is processed together, obscuring individual impacts.. It hides how single clues change the answer and flattens out real uncertainty. To fix this, researchers created LEAP. Instead of processing everything together, LEAP looks at each piece of evidence alone. It measures how likely each clue makes an outcome. Then, a set math model blends those scores with a starting guess to find the final odds. When given the same evidence, LEAP improved most forecast accuracy and calibration scores across models compared to the standard all-at-once approach.

Why it matters

People in finance and sports make decisions based on forecasts. Current methods can hide how evidence affects predictions, leading to unreliable results. LEAP makes forecasts more accurate and transparent, helping users trust and act on them better.

Questions answered