Live dashboard
Market Sentiment Pipeline
Five large-cap tickers. Every weekday after the US close, a scheduled job pulls prices and headlines, Claude scores the news, and a Random Forest calls the next session BUY, SELL or HOLD. Every call is then checked against what the price actually did.
Loading the latest run…
Latest calls
| Ticker | Call | Confidence | Close | News sentiment | Why (the model's one-line reason) |
|---|
Close
Call record
News sentiment
Every call for , as a table
| Session | Call | Confidence | Next-session move | What happened | Result | Source |
|---|
Run history
Every run writes a row here. Rows changed counts inserts plus updates that actually changed a value, so a rerun of a day that already loaded shows 0.
| Started | Trigger | Status | Rows changed | Took | Notes |
|---|
Read this before the numbers
- This is not a trading strategy. It's a data pipeline that happens to produce a signal. There's no position sizing, no transaction costs and no slippage here, so none of this says anything about making money.
- How a call is graded. A call on session D is right if the close on the next session moved the way it said: up more than 0.5% for BUY, down more than 0.5% for SELL, within ±0.5% for HOLD. Those are the same thresholds the model is trained on.
- The baseline. Each hit rate sits next to what you'd score by always calling whichever outcome happened most often over the same sessions. A model that can't beat that isn't adding anything.
- Backtest vs live. Backtest calls were made after the fact by the same code, walking forward one session at a time and training only on data from before each session. News sentiment didn't exist for those sessions, so that feature is 0 in all of them. Live calls are the ones the scheduled job actually made on the day.
- Prices and headlines come from Yahoo Finance through
yfinance. Sentiment comes from Anthropic's Claude API (Haiku). Data is stored in Supabase Postgres, and this page reads it through a cached, read-only endpoint, so it can be a few minutes behind.