On 21 Nov 2024 Adani fell 22.61%. My model didn't know why, and it wasn't allowed to guess.
Description and use
Ghost Candles identifies statistically unlikely stock days using TabPFN forecasts, flags them as ghosts, and produces short narratives, outperforming a bell‑curve baseline.
Problem
Determine how unexpected a stock move was based solely on its historical data, without external news or explanations.
Outcome
Ghost Candles accurately flagged a small number of truly abnormal days (e.g., Adani’s 22.61% drop on 21 Nov 2024) while keeping false positives low, outperforming a simple bell‑curve baseline.
Workflow
- Collect 750 days of historical price data for a stock
- Transform each trading day into a 17‑column feature table (returns, volatility, volume, distance from 50‑day average, etc.)
- Use TabPFN‑3.5 to predict 20‑day ahead quantile ranges for each day
- Slide the prediction window forward to generate 671 forecast days per stock
- Flag days whose actual return lies outside the 98% prediction band as ghosts
- Classify ghosts as Poltergeist (odds 1 in 100 or worse) or Wraith (outside the band edge)
- Generate a short ghost story with Google’s Gemma via Ollama, obeying the rule that no reason can be given
- Render results on a web page with a chart of glowing dots, cards, and an honesty check on the Oracle
Evidence
TabPFN expected 21 November 2024 to land somewhere between −9.92% and +8.98%. It fell 22.61%, on 19.8 times the usual trading volume.
TabPFN makes this cheap. It’s pretrained on millions of made‑up tables and learns a new table just by reading it, so there’s no training and I didn’t tune a single setting.
Adani Enterprises, 671 forecast days: Ghosts flagged (pure luck would give about 7) – TabPFN 4, Bell curve 24.
- Record date: 2026-11-01T11:13:53.860Z