RSI Divergence Backtest on BTC/USDT (2023, hourly, real fees)

Before I could code, I was a chart-reader who circled divergences on printed charts like they were treasure maps. The circles were always beautiful. The account was not. Divergence is the most beloved pattern on TradingView — everyone has circled it, few have defined it well enough to trade. This post is my attempt to turn the prettiest pattern in the book into an honest, rules-based strategy. Strategy Lab #9.

What bullish divergence means

Compare two swing lows:

• Price makes a lower low (worse than the last one).
• But RSI makes a higher low (momentum is losing steam).

The reading: sellers are running out of gas even as price prints new lows — a reversal is likely. It's compelling because it's a two-dimensional argument; it feels smarter than a single indicator level.

How I defined it (and where I simplified)

"Divergence" is famously subjective, so I made it mechanical and ugly:

• Over each 20-bar window, compare the lowest price and lowest RSI to the previous 20-bar window.
• Bullish divergence = price low is lower, RSI low is higher.
• When detected, go long for the next 24 bars, then flat.

This is a deliberate simplification — real traders eyeball swing points and confluence. I'm testing the bare pattern, which is the honest baseline.

Results

At every cost level this underperforms just holding the coin. The pattern isn't "wrong" — it's late and small. By the time the lower-low/higher-low structure is visible, the reversal has often already happened; the entry is confirmation of a move, not the start of one.

The cost bill (short version)

+19.96% → -31.90%. The most striking number here is actually the naive max drawdown of -13.61% — the pattern at least isn't in the market during most crashes. But that defensive posture only helps if the entries earn their keep; on this data they don't.

Buy and hold comparison

What this does NOT prove

• This is one mechanical definition of divergence, one symbol, one year. A filters version (divergence at RSI extremes, higher timeframe confirmation) is a different experiment — and precisely where the TradingView crowd adds their own voodoo. I'm not claiming divergence is worthless; I'm showing what the bare pattern pays.
• The robust lesson: patterns you can't define mechanically can't be backtested — and if you can't backtest it, you can't know if it works. The circles on the chart were always going to be prettier than the P&L.
• Holding for a fixed 24 bars is arbitrary. Exit logic is where divergence trading really lives, and I chose the simplest honest version.

Code

from backtest_base import fetch, backtest_signal, metrics
from strategy import rsi

df = fetch("BTCUSDT", "binance", "2023-01-01", "2023-12-31", "1h")
low, r = df["low"], rsi(df["close"], 14)
now_low, prev_low = low.rolling(20).min(), low.rolling(20).min().shift(20)
now_r,  prev_r  = r.rolling(20).min(),  r.rolling(20).min().shift(20)
bull = ((now_low < prev_low) & (now_r > prev_r)).fillna(False)

signal = hold_after_signal(bull, bars=24)   # long 24 bars after detection
res = backtest_signal(df, signal, cost_per_leg=0.0005,
                      funding_per_bar=0.0001 / 8.0)
print(metrics(res, 8760))

Reproduce it

cd blog-drafts/scripts
python backtest_base.py --strategy rsi_divergence --symbol BTCUSDT --interval 1h \
    --start 2023-01-01 --end 2023-12-31 --fee 0.0005 --funding 0.0000125

Data: Binance public API, hourly OHLCV, 8,735 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.