Prediction Model
How TraderBro identifies, classifies, and tracks analyst predictions.
What is a prediction?
A prediction is a directional call on a specific symbol at a specific point in time, extracted from an analyst's content (tweet, video, article, or transcript).
Each prediction has:
- A symbol (e.g.
NVDA) - A direction:
bullish,bearish, orneutral - A published date — when the analyst made the call
- An optional price target if stated
- A confidence score (0–100) derived from the language used
- An author reasoning excerpt
Extraction pipeline
Content passes through a two-stage LLM pipeline:
-
Triage — determines whether the content contains a genuine directional prediction. Content that is purely commentary, reposting news, or "credit-taking" (claiming a prior call after the fact) is excluded.
-
Extraction — identifies the symbol, direction, price target, confidence, and reasoning from the content.
Credit-taking detection
A common issue in social media analysis is analysts posting "I called this!" after a stock has already moved. TraderBro's triage stage is trained to detect and exclude these retroactive claims. Only forward-looking predictions count.
Prediction series
When an analyst makes repeated calls on the same symbol, these are grouped into a series:
is_series_first: true— the first call in a run; return credit starts hereis_series_last: true— the most recent call in an active series
An analyst who posts "still bullish on NVDA" five times over three months gets credit for one prediction series, not five separate predictions. This prevents artificially inflating prediction count by reaffirming existing positions.
Null fields
Some fields may be null for valid reasons:
| Field | Null means |
|---|---|
stated_price_target | No specific target mentioned in the content |
is_directionally_correct | Return window has not matured yet |
avg_return_3m | Insufficient matured windows to calculate |
accuracy_rate | Fewer than 10 matured predictions |