Quant Notes¶
Key insights discovered during development and study. Each entry captures the date, the question, the finding, and why it matters.
Entry 1 — Individual Factor IR Is Weak by Design¶
Date: 2026-03-02
Question: All four factors (BBIBOLL, STR, VPD, Vol Ratio) showed |IR| in the 0.05–0.12 range. Are these signals too weak to be useful?
Finding: No — this is normal. Practical IR ranges:
| Range | Interpretation |
|---|---|
| |IR| < 0.05 | Noise, not tradeable |
| 0.05 ≤ |IR| < 0.15 | Weak but real — typical for individual factors |
| 0.15 ≤ |IR| < 0.30 | Solid single factor |
| |IR| ≥ 0.50 | Suspicious — likely overfitted or in-sample |
Why it matters: The Fundamental Law of Active Management:
Combining 10 orthogonal factors with IR = 0.10 gives portfolio IR ≈ 0.32. The game is not finding one brilliant factor; it is combining many weak-but-orthogonal signals.
Entry 2 — Anti-Correlation Requires Aligned IC Signs¶
Date: 2026-03-02
Question: BBIBOLL and STR have strongly anti-correlated daily IC (ρ = −0.80). Anti-correlation should be great for diversification — why did equal-weight combination dilute the signal instead?
Finding: Anti-correlation in the IC series means the two factors "take turns being right." But whether that helps depends on direction:
-
Same IC sign + anti-correlated → great. Both predict the same direction. On days factor A is strong, B is weak — but they still point the same way. Mean IC preserved, variance drops, IR rises.
-
Opposite IC sign + anti-correlated → cancellation. BBIBOLL has IC = −0.022, STR has IC = +0.013. When BBIBOLL fires strongly (IC ≈ −0.05), anti-correlation means STR also fires strongly — but in the opposite direction (IC ≈ +0.04). Average: (−0.05 + 0.04)/2 ≈ −0.005. The signals cancel.
Experimental evidence:
- BBIBOLL + Vol Ratio (ρ = 0.02, same IC sign): composite |IR| = 0.136 > 0.122 (best individual). ✅
- BBIBOLL + STR (ρ = −0.80, opposite IC sign): composite IR diluted. ❌
Interview One-Liner
"Anti-correlated ICs reduce tracking error, but only improve IR if the factors agree on direction. Opposite IC signs turn anti-correlation from a hedge into a cancellation."
Entry 3 — Strategy-Based vs. Alpha-Pipeline Backtesting¶
Date: 2026-03-02
Question: Two backtest paths exist. What is the fundamental difference?
Finding: They embody two distinct mental models:
| Dimension | Strategy ("I Trade") | Pipeline ("I Allocate") |
|---|---|---|
| Unit of analysis | Individual trade | Portfolio weight |
| State | Entry price, holding flag | None (recomputed fresh) |
| Decision | Buy/sell this stock? | Allocate across all stocks? |
| Output | Trade list + P&L | Weight matrix + return series |
| Rebalance | Signal-triggered | Calendar-based |
| Typical role | Trader / execution | Researcher / PM |
Why it matters: Most quant interviews and institutional workflows assume the pipeline worldview. The strategy approach is valuable for execution analysis, but the pipeline is closer to how firms like Citadel, Two Sigma, and AQR structure research.
The bridge is WeightBacktester: it accepts pipeline output (weights)
and produces the equity curve that connects "I allocate" to concrete metrics.
Entry 4 — Kelly Criterion Is a Leverage Tool, Not an Allocation Tool¶
Date: 2026-03-02
The most important lesson from Phase 7. See the full writeup in Position Sizing → The Kelly Lesson.
Summary:
- Kelly solves "how much total capital to deploy", not "how to split capital across N stocks"
- Using scalar \(\mu_i / \sigma_i^2\) for cross-sectional allocation ignores correlations and concentrates into low-vol names
- For BBIBOLL (mean-reversion), this systematically under-weights the highest-alpha stocks (which are high-vol by nature)
- Ablation study: signal-weighted sizing cost 0.68 Sharpe vs equal-weight
- Renamed from "Half-Kelly" to "Signal-Weighted" for honesty
Interview One-Liner
"Kelly tells you how much to bet, not how to split the bet. Using scalar Kelly for cross-sectional allocation ignores correlations and concentrates into low-vol names — exactly the wrong thing for a volatility-driven alpha."