article
I Tried to Predict Apple's Stock Price With Linear Regression. Here's Why You Shouldn't.
Ever looked at a stock chart and thought: “how hard can it be, it’s just numbers going up and down”? It was, in fact, very hard.
The setup
I built what looked, on paper, like a proper model:
- 22 technical indicators (RSI, MACD, Bollinger Bands — the whole buffet)
- Lasso regression with feature scaling
- A time-based train/test split, like a responsible adult
The backtest looked great
| Metric | Value |
|---|---|
| R² | 0.78 |
| RMSE | $2.17 |
| Direction Accuracy | 77.5% |
I was mentally planning my yacht.
Then I tested on the future
Not a held-out slice of the past — actual future data the model had never seen, from January–February 2026:
| Metric | Value | Change |
|---|---|---|
| R² | 0.59 | −19% |
| RMSE | $4.83 | +122% |
| Direction Accuracy | 59.1% | −24% |
59.1% direction accuracy — a coin flip gets you 50%. My “proper” model turned out to be slightly better than a coin.
What went wrong
- Overfitting — the model had memorized the past instead of learning something that generalizes.
- Durbin-Watson = 1.40 — the residuals were autocorrelated, a classic sign linear regression’s independence assumption doesn’t hold for time series.
- Adjusted R² was actually 0.51 — a good chunk of that shiny 0.78 backtest number was just noise absorbed by 22 features.
The lesson
Linear regression assumes markets are, well, linear. They’re not — they’re chaos wearing a suit. If you actually want to predict stock prices, the honest next steps are an LSTM (it at least handles time dependencies), XGBoost (it captures non-linear patterns Lasso can’t), or — more realistically — an index fund.