Testing Kronos Candle Forecasts on Chinese Stocks

Summary

The document describes deploying the small Kronos pretrained model on a quantitative research platform to forecast five-minute candlestick data for stocks. Kronos represents OHLCV sequences as hierarchical discrete tokens using a specialized tokenizer, then applies an autoregressive Transformer. The example uses three days of input data to predict the following two days, compares predicted returns with realized returns, and excludes sideways cases from its directional comparison.

The author reports 56.3% directional accuracy across 100 stocks and says the forecasts skewed heavily downward. They also describe a runtime of about one minute for that batch with five sampling passes on a single G0 server. These figures come from a short, narrowly defined example rather than a controlled or out-of-sample evaluation, and the document does not establish profitability or robustness. Fine-tuning was proposed as future work but had not been tested. Deployment requires installing the model dependencies, obtaining the pretrained model and tokenizer, adapting the prediction example, and passing in five-minute data.

Key ideas

• Kronos converts OHLCV sequences into discrete tokens before autoregressive forecasting.
• The example feeds three days of five-minute stock data into the small model to predict the next two days.
• It compares directional returns while excluding sideways outcomes from the accuracy calculation.
• The reported 56.3% accuracy covers 100 stocks in one described evaluation and does not establish profitability.
• The author had not yet tested fine-tuning, leaving its effect unknown.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.