Trading signal performance metrics
Performance comparison across different strategies. Is the signal profitable after realistic costs (6–7 bps) at selective cuts?
The model was trained on data through April 2025. The backtest covers July 2025 to October 2026.
What one trade looks like
Each scored row is a 15-minute candle. The order goes in at the next open, is held for five one-minute candles, and is closed at the close.
Comparison tables
Filter by period. Toggle cost for net edge. All strategies are compared side by side.
| Strategy | Cut | Period | N | Spearman | Gross (bps) | ±SE (bps) | Net (bps) | Winrate |
|---|
Cost sensitivity
How net edge holds as costs rise from 2 to 8 bps, across all strategies.
| Strategy | Gross | Net @ 2 | Net @ 4 | Net @ 5 | Net @ 6 | Net @ 7 | Net @ 8 | Net > 0 at 8 bps? |
|---|
Break-even cost per event equals the gross edge (bps). The last column is "Yes" only when the net edge at 8 bps is positive. Mean edges carry sampling error (see ±SE on the comparison table), so small positive nets are not distinguishable from zero.
Backtest
Causal, capacity-limited simulation on the tradable universe. Cut thresholds are walk-forward (re-estimated monthly from strictly earlier data; A/B are scored against the same prior pool). Fixed $100k account, 2× buying power, ≤10% of buying power and ≤10% of the signal bar’s volume per trade, capital reserved for each 5-minute hold, whole shares, actual trading days, 252-day annualisation. Costs are flat 0–8 bps or scenarios A/B/C computed per trade from the cost model (IBKR $0.005/share with $1 minimum, spreads, slippage, SEC/FINRA fees). Switch return treatment between capped (±50 bps) and raw (rows beyond ±25% removed).
Portfolio statistics
| Cut | Total return | Annualized | Sharpe | Sortino | Calmar | Max DD | Vol (ann.) | End value |
|---|
Trade statistics (net of cost)
| Cut | Trades | Trades/day | Avg trade | Winrate | Avg net (bps) | Avg win | Avg loss | Best | Worst | Profit factor |
|---|
The cumulative-return and drawdown charts overlay all four cuts; the focus cut additionally has rolling sharpe / volatility / return, a monthly heatmap and a daily-return distribution. Equity starts at 100 and tracks a fixed-size account (profits are not reinvested). Daily return = net P&L / prior equity on trading days; Sharpe/Sortino/Calmar use 252 days and the sample standard deviation. Trade stats are per-trade net returns after the selected treatment and cost. Because thresholds come from past data, methods trade different numbers of events: C/D/E select roughly 55–70% of the target count, so total returns are not like-for-like across methods.
Findings
Key performance highlights.
- On the tradable universe, gross edges run ~6.7–23.7 bps at the top 0.1%, ~6.3–16.3 at 0.3%, ~6.4–14.3 at 0.5% and ~4.6–11.8 at 1.0%, depending on method and period. Standard errors are ~2–4 bps at the tight cuts.
- At 6 bps flat cost, all seven methods are net positive at the 0.1% and 0.3% cuts in every period; at 7 bps only F in Jul–Dec 2025 is not. Many of these net edges are only 1–3 standard errors from zero.
- Differences between methods are mostly within noise (the best method is within 2 combined SE of the runner-up in every period/cut cell), so no scorer is shown to be better than the others.
- Jul–Dec 2025 is clearly weaker than 2026 for every method; the 1.0% cut is the most cost-sensitive.
- With realistic per-trade costs (scenarios B and C), the raw-weighted scorers C, D and E fall to or below zero net edge at several cuts, while the rank-based scorers A, B, F and G stay positive at the 0.1% and 0.3% cuts.
- The causal backtest is capacity-limited (about 2–11 long trades a day at the 0.1–0.3% cuts, ~$12k each), so account returns are modest; capped long-only top-0.1% returns at 6 bps range from about 2% (E) to 10.5% (B) over 322 trading days.
- Edges are partly in-sample (the heads were chosen on this data) and costs are modelled, not measured.
Contact
Questions, feedback or ideas about this project? Get in touch.