k-NN Next-Bar Prediction Backtest on BTC/USDT (2023, hourly, real fees)

I watched the machine trade for a while and then closed the tab. The equity curve wasn't the problem — it was smooth for the first hundred hours, drifting up like a strategy that knew something. Then the fee model showed up and the curve became the sound of a wallet being emptied in reverse. This is the eighteenth and final post in the Strategy Lab series, and it's the one with a machine in it. The machine is not the problem. The frequency is. Strategy Lab #18.

What the model does

k-NN with no exotic machinery: for every hourly bar, take the 1,500 previous bars, find the 7 most similar price patterns (nearest neighbors by a few normalized features like returns and RSI), and predict the next bar's direction by majority vote. If the 7 neighbors mostly went up, go long. Retrain the neighbors on a rolling basis — no lookahead, walk-forward.

That's the setup real beginners think is a shortcut: the computer finds the patterns for you. It found 1,821 of them in one year — because hourly patterns are all alike, so the model happily votes on almost every bar.

Results

The naive run — zero costs — returns +36.82% with a Sharpe of 1.17. That's the trap. It looks like the machine found an edge, which is exactly what ML sells you. There is no edge here; there is a model that wins 50-something percent of next-bar bets and, by churning every few hours, converts a coin-flip into a fee catastrophe.

The fee bill

+36.82% → -99.9976%. Not a round trip, not a red year — the account is gone. The final row loses 99.9976% of its value, which is the largest collapse in the entire series, beating even stochastic's -99.96%. The model predicted correctly about half the time; the exchange collected with certainty every single time. 1,821 trades. 3,641 legs. That's the whole story of machine learning on hourly crypto.

The Strategy Lab scoreboard (all 18)

Sixteen BTC strategies and two AAPL daily ones. The two that survive the worst-case 25bp cost model are the two lowest-frequency ones: linear regression (+75.88% at 25bp) and EMA crossover (+51.17% at 25bp). The two machines-with-churn — this one and stochastic's 1,798 trades — are the two worst outcomes. The variable that decided every row of this table was trade count, not indicator quality.

What this does NOT prove

• This is one recipe: rolling k-NN on hourly data with k=7. ML on daily bars, with few trades and a real train/validation split, is an entirely different experiment — the natural next lab, and the only direction the scoreboard points to.
• A single walk-forward year isn't a model evaluation. No out-of-sample validation beyond the walk-forward split was performed.
• The honest conclusion from 18 strategies is boring and worth repeating: indicators and models do not lose money by being wrong. They lose money by being right too often to pay for themselves.

Code

from backtest_base import fetch, backtest_signal, metrics
from sklearn.neighbors import KNeighborsClassifier

df = fetch("BTCUSDT", "binance", "2023-01-01", "2023-12-31", "1h")
feats = df[["ret1", "ret5", "ret20", "rsi14"]].fillna(0).to_numpy()
target = (df["close"].shift(-1) > df["close"]).to_numpy()

signal = np.zeros(len(df), dtype=bool)
model = KNeighborsClassifier(n_neighbors=7)
window = 1500
for i in range(window, len(df) - 1):
    model.fit(feats[i-window:i], target[i-window:i])
    signal[i] = model.predict(feats[i:i+1])[0]

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 knn --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.