When Factors Stop Working: How Regime-Aware Models Reduce Drawdowns in Indian Equities
Most factor models fail not because signals disappear, but because they are trusted for too long. This note explains how regime-aware factor gating, conditional growth, and governance caps were implemented in RevDog V3 to manage drawdowns in Indian equities.
When Factors Stop Working, Most Models Don’t Notice
Most quantitative equity models do not lose money because their factors stop working.
They lose money because portfolio managers continue trusting those factors after their effectiveness has already deteriorated.
This problem is subtle, persistent, and structural. Backtests often hide it. Live capital exposes it.
RevDog V3 was rebuilt to address exactly this failure mode: knowing when a factor should stop being trusted.
1. Regime-Aware Factor Weighting: Knowing When to Listen
What Walk-Forward Validation Revealed
Using rolling walk-forward validation across market regimes, we evaluated how commonly used factors behaved in normal markets versus stress periods.
The results were uncomfortable—but consistent.
| Factor | Normal Markets | Stress Periods | Interpretation |
|---|---|---|---|
| ROE | Positive IC | Negative IC | Hurts during drawdowns |
| ROCE | Positive IC | Negative IC | Fails when protection matters |
| OPM | Marginal | Strongly negative | Worst stress performer |
| Promoter Holding | Positive | Stronger | Defensive in stress |
| Promoter Change | Weak | Very strong | Crisis-period alpha |
The insight:
High-quality, liquid stocks (high ROE/ROCE) are often sold first during risk-off phases.
Promoter-backed companies, with concentrated ownership and skin in the game, are defended.
What works in bull markets can actively hurt during drawdowns.
A useful analogy is defensive driving: speed matters on a clear road; survival matters in fog. Same vehicle, different priorities.
How This Is Implemented
The system operates in three deterministic modes:
| System Mode | Market Context | Equity Exposure | Cash | Objective |
|---|---|---|---|---|
| SURVIVAL_MODE | Market stress | ~37% | ~63% | Capital preservation |
| ALPHA_MODE_CAUTIOUS | Uncertain | ~67% | ~33% | Risk-controlled participation |
| ALPHA_MODE | Bullish | ~92% | ~8% | Full factor expression |
Regime classification is driven by five independent macro indicators:
- Small-cap relative performance
- India VIX percentile
- USD/INR trend
- Gold vs equity performance
- Bank NIFTY relative strength
Defense is triggered faster than offense. Two stress signals are sufficient to turn defensive; three positive signals are required to re-risk. The asymmetry reflects a simple institutional reality: missing upside is survivable; large drawdowns are not.
2. Conditional Growth in Survival Mode
The Question
If the system is in survival mode, should growth be ignored entirely—or can it still add value?
What the Data Showed
Walk-forward testing compared pure survival factors against blended models.
| Configuration | Change in IC |
|---|---|
| Survival only | Defensive, no alpha |
| Growth only | Alpha exists, unstable |
| Survival + 25% growth | Meaningful improvement |
| Survival + 50% growth | Higher but unstable |
The conclusion: growth adds value only after survival is ensured.
Implementation Choice
Instead of binary gating, V3.1 applies discounted growth in stress:
final_score = survival_score + 0.25 × growth_scoreAt this weight:
- Survival remains dominant
- Growth improves discrimination among defensives
- Weak survival stocks cannot “rescue” themselves via growth optics
The improvement is incremental, not transformational—but statistically real.
3. Governance Overrides: Capping Risk, Not Forcing Exits
Promoter Pledge (>80%)
High promoter pledge creates asymmetric downside. Falling prices trigger margin calls, which force selling, often accelerating the decline.
Some of India’s most severe equity collapses followed this pattern.
Rule:
If promoter pledge exceeds 80% of holding, the stock’s maximum conviction is capped at low levels.
Below 50% is common and manageable.
Between 50–80% requires monitoring.
Above 80% represents structural fragility.
The rule does not exclude the stock. It limits how much confidence the system can express.
Cash Burn (Non-Financials)
Sustained negative operating cash flow over multiple years often signals deeper structural problems.
Rule:
- If operating cash flow is negative for two consecutive years, conviction is capped.
- Financial services are excluded, as negative OCF during loan book expansion is normal.
This catches working-capital stress early without penalizing legitimate lending businesses.
Governance Philosophy
Governance rules do not remove discretion.
They impose risk ceilings, not opinions.
Every flag is visible, logged, and auditable. Nothing is hidden inside composite scores.
4. Walk-Forward Validation: Why This Matters
Backtests can be optimized indefinitely. Walk-forward validation is less forgiving.
RevDog V3 uses:
- 15-month training windows
- 6-month test windows
- Quarterly roll-forward
- 45-day disclosure lag
Factors are judged by:
- Information Coefficient (IC)
- Hit rate
- Sign stability
A factor that flips sign in stress is not neutral—it is dangerous.
5. Factors That Were Removed
Several popular signals failed basic out-of-sample tests:
| Factor | Reason for Removal |
|---|---|
| FII/DII holdings | Stale and already priced |
| FII/DII changes | Noisy, unstable |
| OCF/PAT ratio | Dilutive |
| Working capital days | Negative IC |
Institutional flow data, when disclosed quarterly, often reflects what has already happened—not what will.
6. What Did Not Change
Not everything needed fixing.
- Eligibility filters remain intact
- Sector-relative normalization still works
- Temporal smoothing is retained
- Position and sector caps remain unchanged
The system changed when factors are trusted—not the foundational risk controls.
Appendix: Interpreting Risk–Return Behavior Across Regimes
| Dimension | V2 (Pre-Regime) | V3 (Regime-Aware) | Design Intent |
|---|---|---|---|
| Factor Usage | Static across cycles | Gated by regime | Avoid factor inversion |
| Growth Exposure in Stress | Full | Discounted | Preserve selectivity |
| Capital Deployment | Always invested | Dynamic (defensive in stress) | Drawdown control |
| Governance Signals | Averaged into score | Conviction caps applied | Prevent false confidence |
| Response to Drawdowns | Reactive | Pre-emptive | Reduce loss severity |
| Whipsaw Protection | Limited | Smoothed, asymmetric | Avoid premature risk-on |
Interpretation:
This system is engineered to remain deployable across market regimes rather than to optimize performance in any single regime.
The observed differences reflect intentional trade-offs: reducing drawdown severity and signal inversion during stress at the cost of slower re-engagement in sharp recoveries.
The objective is stability of interpretation when market behavior changes—not maximizing headline returns in favorable windows.