Week 3: Cluster Detection Holds, Regime Shift

The Cluster Fix Held

The biggest question coming into week 3: would the cluster detection actually work under pressure, or did we just get lucky patching it after the fact?

Answer: it held. On Day 17, three correlated mean reversion setups appeared on BTC within 70 minutes — nearly identical conditions to the ETH clustering incident from week 2. The first signal executed. The second was blocked (same underlying, cooldown active). The third was also blocked. Only one position opened.

BTC dropped 2.8% during that window. The single capped position lost $23.10. Without cluster detection, it would have been three times that. The fix earned its keep.

Regime Flip Mid-Week

Wednesday morning the regime classifier flipped from "ranging" to "trending" for the first time since launch. ADX crossed 25, Bollinger Bands expanded, volume spiked on the breakout candle. The volume profile input we added this week confirmed the signal — high-volume node cleared, no nearby resistance.

The system responded correctly: mean reversion strategies were suspended. Momentum strategies were upweighted. Two momentum trades executed in the trend window — both profitable.

The regime flipped back to ranging by Friday. The system caught both transitions cleanly. This is the adaptive behavior we designed for — and the first real test of it.

Technical Work This Week

• Volume profile added to regime classifier. Third input alongside ADX and Bollinger Band width. Reduces false regime classifications during low-volume consolidation periods. Early result: fewer false "trending" signals on thin markets.
• Inter-strategy correlation matrix live. Quantifies overlap between all 9 active strategies by underlying asset, entry conditions, and signal correlation (rolling 30-day). Four strategy pairs have correlation >0.6 — flagged for review in week 4.
• Equity curve tracker deployed. Cumulative P&L chart, rolling 7-day Sharpe approximation, and max drawdown tracking are now live on the dashboard. First Sharpe reading (7d): 0.41 — too early to be meaningful, but the tracking is running.
• Polymarket resolution classifier v2 in production. Scanned 34 markets this week. Rejected 6 that v1 would have flagged as tradeable due to ambiguous resolution wording.

What Broke (Because Something Always Breaks)

Issue: Regime classifier fired a false "trending" signal on Day 16 — low ADX (21) but Bollinger Bands happened to be wide from a single high-volatility candle. Volume profile data wasn't yet live to correct it.

Result: Momentum strategy activated briefly, found no qualifying setups, and deactivated. No trades taken, no capital at risk. But the false classification held for 47 minutes.

Fix: Volume profile input went live the same day. The Day 18 regime transition used all three inputs and was a clean, unambiguous flip.

Issue: Dashboard equity chart used local time instead of UTC for timestamps. Looked fine for us (EST), broken for anyone in a different timezone.

Fix: Normalized all chart timestamps to UTC with explicit label. Minor, but worth documenting.

The High-Correlation Strategy Pairs

The correlation matrix surfaced something we need to address. Four strategy pairs are showing >0.6 rolling correlation — meaning when one fires, the other is likely to fire within the same session:

• Mean reversion + Bollinger Bounce (ETH): r = 0.74
• RSI oversold + Fear/Greed contrarian: r = 0.68
• Momentum breakout + Volume surge: r = 0.71
• Polymarket arb + On-chain sentiment: r = 0.63

These aren't bugs — they're expected. Related strategies share signal inputs. But they mean our "9 active strategies" is not 9 independent signals. It's closer to 5-6 independent bets dressed as 9. We're evaluating whether to merge the most correlated pairs or accept the overlap and manage it through position sizing.

What We Learned

1. Regime classification is the most important variable in the system. Not the strategies. Not the LLM confidence scores. The regime. Running momentum in a ranging market is how you bleed. Running mean reversion in a trend is how you get stopped out repeatedly. The regime classifier made the right call twice this week.

2. “9 strategies” is a marketing number, not a risk number. The correlation matrix shows we have ~5-6 independent signal sources. That's still diversification, but it's honest diversification. We're updating how we describe the system internally.

3. The equity curve changes how you think about the system. Having cumulative P&L visible in real-time shifts the psychology of the whole operation. Losses feel different when you can see the drawdown percentage live. That's a feature — it creates discipline around the circuit breakers.

Week 4 Focus

• Decide: merge high-correlation strategy pairs or accept overlap and manage through position sizing
• Run full backtest of the 3-input regime classifier on 2024-2025 historical data — validate the volume profile addition wasn't overfit to recent conditions
• Stress-test the momentum strategy under the current trending regime if it persists
• Publish equity curve chart publicly on the challenge page (currently dashboard-only)

The Bottom Line

Three weeks in. Cumulative paper P&L: +$244.40 (+2.44%). Max drawdown: −0.9%. Zero days of downtime. One overtrading incident (week 2), caught and fixed. One false regime classification (week 3), corrected before any capital deployed.

The system is behaving like it should in the early phase: conservative, adaptive, loud about its own failures. We're 23% of the way through. Compound effects start showing up in weeks 6-8 if the patterns hold.

Week 4 report publishes Monday April 13. Same format — wins, losses, and whatever breaks.