Prediction Strategies

Single-prompt categorical prediction (very_negative to very_positive). The simplest approach: give the model a press release and ask it to classify the stock impact into one of 7 categories.

Step-by-step reasoning before categorical prediction. Forces the model to analyze key results, statistical significance, clinical meaningfulness, safety signals, regulatory implications, and market context before making a prediction.

Single-prompt numeric scoring on 0-10 scale. Instead of categorical labels, asks the model to output a calibrated numeric score where 0 is catastrophic and 10 is exceptional.

5-step agent pipeline: Extract Facts > Web Search > Clinical Assessment > Market Dynamics > Final Prediction. Uses specialized prompts at each step and aggregates findings for a more informed prediction.

5-step agent pipeline producing numeric 0-10 scores. Same multi-step architecture as the categorical agent but outputs calibrated numeric predictions instead of categories.

Predict percent change via ridge regression using market cap, momentum, and LLM-quantified qualitative features extracted from the press release and clinical trial data. The LLM acts as a feature extractor, not a direct predictor.

Ridge regression with market cap, momentum, PR/trial LLM features, and pre-event web search features. Alpha tuned via inner cross-validation. Combines both LLM feature extraction and real-time market context.