Experiment · live
Can machine learning beat naive baselines at predicting whether NVDA closes up or down the next day?
An experiment, not advice. Nothing on this page is a recommendation to buy or sell anything. The calls below are published so they can be scored, not so they can be followed.
The stock
The chart is provided by TradingView and shows their data. The models on this page do not use it; they work from daily closing data only.
The latest calls
Live, since the log started
Every call is written to a log before the market opens and scored after it closes. The log can only be added to; see the Method page for how that is enforced.
Tested on history
Before going live, every baseline and model was tested on past data, always predicting a day using only the days before it.
sessions tested, 2004 to 2026
of those days were up: the bar to beat
models beat always calling up
| Predictor | Accuracy | Simulated return per year after 5 basis points per trade |
Return vs buy-and-hold |
|---|---|---|---|
| Always up Calls up every day. Following it is buy-and-hold. |
52.4% (51.1% to 53.7%) |
37.8% (13.0% to 66.3%) |
benchmark |
| Persistence Calls whatever the last session did. |
48.7% (47.4% to 50.0%) |
2.7% (-10.3% to 16.8%) |
-35.1 pts (-56.0 to -16.9) |
| 200-day average rule Up when the price is above its 200-day average. |
52.3% (51.0% to 53.6%) |
39.7% (20.1% to 63.9%) |
+1.9 pts (-15.1 to +18.6) |
| Logistic regression A weighted sum of the features. |
50.8% (49.5% to 52.1%) |
24.9% (7.7% to 45.2%) |
-12.9 pts (-31.0 to +3.4) |
| XGBoost Boosted decision trees on the features. |
50.7% (49.3% to 52.0%) |
22.3% (6.5% to 40.1%) |
-15.5 pts (-35.0 to +1.2) |
| LSTM A neural network reading the last 20 sessions. |
51.6% (50.2% to 52.9%) |
27.0% (10.2% to 46.6%) |
-10.8 pts (-30.0 to +7.2) |
Ranges in brackets are 95% intervals. Simulated returns are summary figures from applying each rule to history; they are not real results.
How it was done
Each day after the close, every predictor makes one call: will the next session close higher than this one? “Higher” uses the closing price adjusted for dividends and splits, and a flat day counts as not up.
Three rules that need no learning, and that any model has to beat:
Each model’s settings were fixed before it was tested and were not adjusted afterwards to improve the result.
Twenty numbers describing recent price and volume behaviour, all computed from data up to that day’s close: returns over the last 1 to 21 sessions, how volatile the last week and month were, two standard momentum indicators (RSI and MACD), how far the price sits from its 10, 50 and 200-day averages, volume compared with its recent average, the day’s high-to-low range, and the day of the week.
Testing is walk-forward. A model is trained on the first five years, makes its calls for the next 21 sessions, is retrained with those sessions added, and so on to the present. No call is ever made by a model that has seen that day or any day after it. Automated tests replace all later data with nonsense and check that earlier calls do not change.
The simulated strategy holds the stock on days called up and holds cash on days called down. Each change of position is charged 5 basis points (0.05%) of the amount traded. Cash earns nothing in the simulation.
Page data last built 2026-11-19 05:32 UTC.