Linear Regression Channel Backtest on AAPL (2018–2025, daily, real fees)

Fifteen strategies into this lab, fees have eaten everything: hourly crypto churns, and every churn pays the exchange. Then I ran the least glamorous thing on my screen — a straight line fitted through 60 days of closing prices — and it became the first strategy that survives a realistic cost model. No fancy entries, no multi-timeframe filters, no machine learning. Just OLS, on daily AAPL, for eight years. Strategy Lab #17.

What a linear regression channel is

For each bar, fit a straight line through the last 60 closes using ordinary least squares:

close_t ≈ a + b·t

The fitted line is the trend of the moment; the spread of residuals tells you how far price has wandered from it. My rule is the simple channel reading:

• Long while close > the 60-day regression line.

That's the entire strategy — 106 trades over eight years, about 13 a year. The sample is AAPL daily (2018–2025), so there's no funding cost at all; the only fees are the two taker legs per round trip.

Results

The honest-cost line is +199.09% with a -23.34% max drawdown. Now compare that to just holding the stock: buy & hold made +576.91% but lost -38.52% at its worst — and that drawdown was the 2022 bear, which people actually lived through. The regression line gets you half the return with half the pain and a nearly identical Sharpe (1.02 vs 1.12). That's the trade the chart is offering: less money, more sleep.

Every cost scenario is positive

| Scenario | Cost/leg | Total return | MaxDD | Sharpe | Trades | |---|---|---:|---:|---:|---:|---:| | naive (zero cost) | 0.00% | +232.54% | -22.96% | 1.11 | 106 | | taker fee 0.05%/leg | 0.05% | +199.09% | -23.34% | 1.02 | 106 | | + funding 0.01%/8h | 0.05% | +199.09% | -23.34% | 1.02 | 106 | | + slippage 10bp/leg | 0.15% | +141.90% | -24.77% | 0.84 | 106 | | + slippage 25bp/leg | 0.30% | +75.88% | -34.53% | 0.58 | 106 |

Every row is green. Not because the indicator is magic, but because it trades 13 times a year instead of 700. Compare with OBV (#16): identical logic (series above its own trendline), 106 trades vs 734, +75.88% vs -98.10% at the same 25bp cost. The difference is frequency, not foresight. This is the single most important number in the whole series, and it's the trade count.

Buy and hold comparison

What this does NOT prove

• A fitted line is not a prediction. The regression channel describes the past 60 days; it has no opinion about tomorrow. What made it survive is that it stays in stocks that are above their own trend and leaves before the drawdown deepens — a trend filter, not a crystal ball.
• One very strong stock, eight years, daily bars. AAPL has an excellent 2018–2025; the same channel on an index or a different decade will differ. What survives a regime change is the low-frequency principle, not the 60-day parameter.
• This is the setup for the last post in the lab (#18): statistical signals on daily data are the one branch that holds up under costs — so the next step, machine learning, should start here, not on hourly crypto.

Code

from backtest_base import fetch, backtest_signal, metrics
import numpy as np

df = fetch("AAPL", "yahoo", "2018-01-01", "2025-12-31", "1d")
close = df["close"]
t = np.arange(len(close))

def reg(ts):                       # OLS fit over a 60-bar window
    x = ts[-60:]
    slope, intercept = np.polyfit(np.arange(60), x, 1)
    return intercept + slope * 59  # line value at the last bar

line = close.rolling(60).apply(reg, raw=True)
signal = (close > line).fillna(False)
res = backtest_signal(df, signal, cost_per_leg=0.0005, funding_per_bar=0.0)
print(metrics(res, 365))

Reproduce it

cd blog-drafts/scripts
python backtest_base.py --strategy linear_reg --symbol AAPL --interval 1d \
    --start 2018-01-01 --end 2025-12-31 --fee 0.0005 --funding 0

Data: Yahoo Finance, daily OHLCV, 2,010 bars. The tables above reproduce exactly from this command.

This is a backtest on historical data, not investment advice. Past performance does not predict future results.