How the AI stock forecasts work

Short version: once a day, after the US market closes, an open-source AI model reads each stock's recent daily prices and predicts the next session's close. Then, a day later, we check how it did and publish the result.

1. The data

For every covered stock and ETF we download about two years of daily bars — open, high, low, close and volume — from Yahoo Finance's public chart data, shortly after the US close. Prices are daily, not real-time, and can occasionally contain errors from the source. Only finished trading sessions are used.

2. The model

Forecasts come from Kronos-small, an open-source foundation model for financial time series released under the MIT license. Kronos was pre-trained on billions of candlestick records from dozens of exchanges; it turns price bars into tokens, much like a language model turns text into tokens, and predicts what comes next.

For each stock it reads up to the last 450 trading days of daily open, high, low, close and volume. The model is "zero-shot": it is not fine-tuned per stock, and it sees nothing except that price and volume history — no news, earnings, fundamentals or macro data.

3. The forecast and the range

Kronos samples its predictions, so running it twice gives two slightly different answers. We run it 16 times for every stock. The forecast you see is the median of those 16 closing prices, and the likely range runs from the 10th to the 90th percentile — roughly where the model expects the close to land 8 times out of 10. A wide range means the model is uncertain.

4. The track record

The next evening, each forecast is compared with the actual close. For every stock we publish the last forecasts with the prior close, the predicted close, the real close, the error, and whether the predicted direction (up or down from the prior close) was right.

5. What to expect from it

Day-to-day stock moves are dominated by news and noise, so no model — human or AI — predicts them reliably. A direction hit rate a little above or below 50% is normal. The value of this site is seeing what a modern time-series AI "thinks", how uncertain it is, and how it performs in public over time. It is not a trading signal.

Schedule

• Data and forecasts update once per US trading day, after the close.
• Forecasts are for the next regular trading session (weekends and NYSE holidays are skipped).
• If an update fails, the last forecast is marked as pending rather than reused.