Up or Down About

Strategy backtest: Amazon

Each model's prediction at today's close sets tomorrow's position (long or short). Returns are daily log returns over the test period (7 Jun 2021 to 24 Sep 2026), net of costs charged on every position change.

Verdict

1 of 7 strategies beat buy & hold on Sharpe ratio

With hindsight, the best on the test period is SVM (RBF) at 0.50 against 0.24 for buy & hold. Drag the cost slider to see how quickly turnover erodes any edge.

Risk and return

Annualised with 252 trading days. Sharpe and Sortino assume a zero risk-free rate; Calmar = annual return รท max drawdown.

StrategyTotal returnAnnual returnVolatilitySharpeSortinoMax drawdownCalmarHit rateTradesCosts
SVM (RBF) +157.4% +19.6% 35.7% 0.50 0.72 45.5% 0.43 51.0% 252 25.15%
Random Forest +30.0% +5.1% 35.8% 0.14 0.20 52.6% 0.10 50.6% 171 17.05%
Extra Trees +24.4% +4.2% 35.8% 0.12 0.16 58.9% 0.07 50.8% 59 5.85%
Logistic Regression +9.9% +1.8% 35.8% 0.05 0.07 77.3% 0.02 50.1% 240 23.95%
Linear SVC -6.0% -1.2% 35.8% -0.03 -0.05 78.1% -0.01 49.5% 236 23.55%
Bagging (KNN) -6.6% -1.3% 35.8% -0.04 -0.05 46.4% -0.03 48.2% 541 54.05%
Gradient Boosting -28.9% -6.2% 35.8% -0.18 -0.25 69.7% -0.09 49.8% 429 42.85%
Buy & hold +56.1% +8.8% 35.8% 0.24 0.34 56.1% 0.16 50.9% 1 0.05%

Costs and risk

Would it survive real trading, and was the return worth the risk?

Takeaway With free trading, 2 of 7 strategies beat buy & hold on Sharpe; at 50 bp, none do. SVM (RBF) changed position 252 times in 1,332 trading days.

Takeaway In long/short mode every strategy is always fully invested (long or short), so they all carry the stock's volatility. Only the return differs.

Inside one strategy

Change the model with the "Detail model" selector