Why Macro Factors Beat Multi-Factor ML for Signal Distribution

[michaeljiangmingfeng-debug]

A question we get often: Why does QuantToGo use macro-factor strategies instead of multi-factor machine learning models?

The short answer: macro factors are the only type of quantitative alpha that works well as an externally distributed signal.

Here's why:

The Problem with ML-Based Signals

Multi-factor ML models (random forests on 200 alpha factors, deep learning on order book data, etc.) have a fundamental problem when used as signal sources:

1. Black box — Even the creator often can't explain why a specific trade was triggered. If users can't understand the signal, they won't trust it enough to act on it.
2. 2. Overfitting risk is invisible — A beautiful backtest curve can be entirely data-mined. Without understanding the causal mechanism, there's no way to distinguish a robust factor from a statistical coincidence.
3. 3. Regime sensitivity — ML models trained on 2015-2024 data may completely fail in 2025 if the market regime shifts. Macro factors, by contrast, are driven by economic fundamentals that persist across regimes.
4. 4. Capacity destruction — Many ML-based alpha signals have tiny capacity. A few hundred followers and the signal is arbitraged away.

Why Macro Factors Work Better

Our 8 strategies are all driven by macro-economic factors with clear causal chains:

Factor Type	Example	Why It Persists
FX-Equity Correlation	CNH depreciation → CSI 300 follows in 1-2 weeks	Capital flow mechanics between offshore and onshore markets
Sentiment Extremes	VIX spike > 35 → mean reversion within weeks	Markets systematically over-price fear during panics (behavioral bias)
Liquidity Rotation	Large-cap/small-cap spread widens → rotation signal	Institutional fund flows follow predictable cycles
Put/Call Ratio	Extreme retail put buying → contrarian buy signal	Retail option traders are consistently wrong at extremes

Each of these has:

• Clear economic logic — you can explain it in one sentence
• • Decades of persistence — driven by structural market mechanics or human behavior
• • Medium frequency — daily to monthly signals, giving users time to execute
• • Index-level targets — liquid ETFs/futures, no individual stock picking needed

The Signal Source Test

We think any signal source should pass this test:

Can a non-programmer, non-quant individual investor receive the signal, understand why it was generated, and execute it in their own brokerage account within a reasonable timeframe?

Macro factor strategies pass this test. Most ML-based strategies don't.

This is also why we open-source all forward-tracked performance data. Every signal is a git commit — timestamped, immutable, including all losses. If the factor logic is sound, the track record speaks for itself.

---

Curious what others think. Are there ML-based signal types that could work well for external distribution? What would they need to look like?