Ignacio Such

Founder The-Bou | Private Investor

[Ignacio Such @ignacio_such]

Full deep dive (3,800 words):

• Architecture explained
• • V1→V2→V3 evolution
• • What I’d do differently
• • Why defensive > aggressive for compounding
https://t.co/T7KzHCQyQI

If you’re building systematic trading systems and value honesty over hype, subscribe.

Next post: what V3 actually learned.

[Ignacio Such @ignacio_such]

Early V3 signals (combos 2-5 of 84):

CAGR: +4% → +17% → +17% → +22% ⬆️
Sharpe: -1.84 → +0.12 → +0.31 → +0.42 ⬆️
Bear fraction in buffer: 24% → 33% → 31% ✓

The model is learning. Each combo better than the last.

Full training: ~13 days remaining.

[Ignacio Such @ignacio_such]

What’s coming:

✓ V3 baseline complete (~13 days)
✓ Stress framework V3 vs V1 deterministic
✓ If passes 4/6 criteria → @Darwinex_Zero deployment
✓ Public verifiable track record begins

This newsletter documents everything. Drawdowns included.

[Ignacio Such @ignacio_such]

V3 (currently training) carries V2 + 5 additions:

• CAAN attention (cross-asset)
• • Turnover penalty (fixes V2 fee problem)
• • Calendar features (FOMC, quarter-end, earnings)
• • Vol targeting (institutional risk mgmt)
• • Hard MaxDD circuit breaker @ 25% + HAR-RV vol forecast (Corsi 2009)

[Ignacio Such @ignacio_such]

I built a defensive RL trading system. During the Lehman 2008 replay, it lost 22%. The S&P lost 55%. The story of 18 months iterating on V1 → V2 → V3 below 🧵

[Ignacio Such @ignacio_such]

V2 training revealed an unexpected problem:

Excessive turnover. ~20 trades/day. $40K in fees per simulated episode.

The Ulcer penalty alone didn’t reduce churn—the agent was rebalancing constantly to maintain its risk profile.

Lesson: reward shaping has interaction effects.

[Ignacio Such @ignacio_such]

V2 attempted to fix slow-bear weakness. I consulted @GoogleDeepMind’s Gemini Deep Research.

Three modifications selected from peer-reviewed lit:
1. Ulcer Index reward penalty (Martin 1989)
2. 2. HMM regime injection + stratified replay (anti-amnesia)
3. 3. MDD auxiliary prediction head (DeepScalper, Sun 2022)

[Ignacio Such @ignacio_such]

V1 results (in-sample, with stress testing):

• COVID 2020: -21% vs $SPY -34% ✓
• • Volmageddon 2018: -9% vs $SPY -10%
• • CCAR 2025 GMS (-50% shock): -2% (synthetic stress)
• • Bear 2022: -16% vs $SPY -17% ⚠️ (weak)
Fast crises: excellent. Slow grinding bears: mediocre.

[Ignacio Such @ignacio_such]

Why distributional RL?

The model doesn't predict an *expected* return.
It predicts a *distribution* of returns.

This makes it natively aware of tail risk - the worst-case scenarios get weighted in decision-making.

Standard RL maximizes mean. Distributional RL respects variance.

[Ignacio Such @ignacio_such]

The stack:

• TQC (Truncated Quantile Critics) - distributional RL
• • 25 ETFs: equities, bonds, gold, FX safe havens ($SPY $TLT $GLD $FXF $FXY)
• • CPCV validation (López de Prado 2018) - 15-28 unique backtest paths
• • Stress framework: CCAR, vine copulas, reverse stress (Glasserman 2012)

[Ignacio Such @ignacio_such]

The benchmark I set:

Bridgewater All Weather lost ~22% in Lehman. Most mutual funds lost 50%. SPY lost 55%.

That 22% number became my target.

(Spoiler: V1 matched it. V3 is training to beat it.)

[Ignacio Such @ignacio_such]

Most retail quants try to beat the market. I tried to *survive* it. Asymmetric math:
• 50% loss needs +100% to recover
• • 30% loss needs +43%
• • 20% loss needs +25%
The best defensive systems don't chase alpha. They preserve compounding.

[Ignacio Such @ignacio_such]

@pobremillenial Se espera subida de tipos para 2028?

[Ignacio Such @ignacio_such]

@TrendSpider Currently all smoke, good tech indeed, but such rating?

[Ignacio Such @ignacio_such]

Bridgewater All Weather lost 22% during the 2008 crash.

A retail system I built — running on a single RTX 3060 — matched that number on the same in-sample replay.

Tomorrow 14:00 CET: how I did it, what's broken, what I'm fixing in V3.

The engineering thread, not the pitch.

[Ignacio Such @ignacio_such]

Stop optimizing for Sharpe.

Sharpe penalizes upside volatility equally with downside.

Your "perfect Sharpe 2" strategy might be one that never lets you win.

Defensive systems should optimize for Calmar (return / max drawdown) or Ulcer Performance Index.

The metric you choose IS the strategy.

[Ignacio Such @ignacio_such]

Oct 9, 2007: S&P 500 hits all-time high.

Mar 9, 2009: S&P 500 down 55%.

A defensive RL system I built lost 22% in that same window.

I wrote 3,800 words about how — and what I learned trying to fix the slow-bear weakness 14 years later.

https://t.co/jlVmBY6374

[Ignacio Such @ignacio_such]

@pobremillenial @XMihura Como valenciano apruebo este tweet