Mikhail Shardin conducted an experiment: can a language model predict prices if charts are described in text?

The Idea

Instead of raw quotes, the model received natural language descriptions: price rising strongly, volume increasing, near resistance.

The DistilBERT model was trained to predict next-day price increases.

Results

Tested on 200+ Moscow Exchange stocks:

  • Average AUC: 0.53 (slightly better than random)
  • Best performers: AFLT (0.72), RTSB (0.70), PIKK (0.70)
  • Worst performers: PLZL (0.33), VJGZP (0.33)

For trading purposes the result is weak, but the model picked up patterns without direct access to numbers – that alone is interesting.

Technology

Python + PyTorch + Hugging Face + Docker. Walk-forward validation, vectorized processing via pandas. The entire process is reproducible.

Code on GitHub: github.com/empenoso/llm-stock-market-predictor

Source: Habr Author: Mikhail Shardin