TRADING : UN QUANT EXPLIQUE COMMENT BATTRE LE MARCHÉ | MONEYTALK #33

Finance-focused summary of the subtitles (MoneyTalk #33)

Core theme / claims

• Retail trading is often undermined by:
  • Backtest illusions
  • Overfitting
  • Survivorship bias / shiny-object bias
• Profitability is framed as coming from rigorous data, risk limits, and system-level execution—not “miracle” strategies.
• Institutional/quant investing is described as:
  • Probability modeling + infrastructure + risk constraints
  • With the key caveat that edges degrade over time.

---

Key instruments / tickers / assets / sectors mentioned

Equity benchmarks

• S&P 500
• Nasdaq

Sectors / indices

• Semiconductor sector / SOX
• Nvidia (referenced in terms of valuation risk and infrastructure exposure)
• S&P 500 constituents (general discussion)

Commodities / metals

• Gold
• Silver

Crypto / exchanges and venues

• Bitcoin (referenced around the 80,000 price level)
• Binance (market disruption cited)
• BNB (mentioned as the source of a market event on Oct 10)
• Hyperliquid (data mention)

Options / futures

• Options (pricing/risk models, delta-hedging / “deltaing”)
• Futures indices (used as hedges; deltaing mentioned)

Other entities / historical examples

• LTCM (failure despite “models”)
• JP Morgan / Citadel (mentioned as Citadelle/Citadel in context of option traders and risk limits)

---

Numbers & explicit metrics cited

Backtesting / signal existence & randomness

• FVG experiment (Fair Value Gap / “FVG rule”):
  • 67% chance of returning to the zone under the FVG rule
  • 92% chance under random mode
  • Conclusion: randomness beats the FVG rule, suggesting the “structure” may be largely random
• Transferability of backtest performance:
  • Poorly done backtest: may realize only ~25–30% of backtested performance over 1 year
  • Well done backtest: could transfer ~80–85%
  • Claim: you should never expect 100% because markets evolve

Probability / alpha / benchmarks

• Alpha example:
  • If a benchmark (e.g., S&P 500) gained ~18% and you gained 20%
  • Then ~2% alpha
• Probability framing:
  • 5–10% probability is described as “huge” in quant terms (model-based, not story-based)

Risk limits & drawdown tolerance

• Max acceptable losses / hit levels:
  • Depending on strategy: below ~20%
  • Retail: “flexible up to ~20%,” but “they will never exceed 5–6%” (as stated)
• Capital/risk framing example:
  • If capital is $1000, “mustn’t go below $800” (~20% max drawdown rule of thumb)
• Quant fund constraint example:
  • Citadel option trader limit cited: ~7% maximum annual return, or risk of being fired (presented as strict internal limits)

Trading/edge degradation

• Edge death / maintenance:
  • “Impossible to maintain an edge for 10 years”
  • Allow roughly 20–25% degradation; if performance drops that much, “your edge is gone”
• Regime change:
  • Edge/regime shift is used to justify diversifying strategies

Market microstructure / execution

• Human vs machine processing:
  • Human visual processing example: ~25 images/sec (with “up to 30” mentioned)
  • Market open information density: ~700 out of a second (tape data example)
  • Retail order entry: typically 10–13 seconds
• Machines:
  • Continuously operating and able to process many assets simultaneously

Crypto events / liquidation sizing (approx.)

• Binance event referenced:
  • Oct 10 cited as an “initiated lead”
  • VPIN model referenced; market “turned,” orders blocked
  • “It was still, I think, 1.5 billion” (approx; possibly more)
• Later liquidation:
  • “Yet another liquidation… on the rise of Bitcoin to 80,000”
  • Size not clearly quantified beyond the prior ~1.5B+ context

Infrastructure / AI cost forecast

• LLM/API costs:
  • “Each GPT will cost more than $20”
  • By end of 2027, it “will be over” (i.e., prohibitively expensive)
• Infrastructure example:
  • “€3000 for a small infrastructure at home”
• Tooling:
  • Building custom AI models requires infrastructure

---

Methodologies / frameworks explicitly shared (step-by-step style)

1) Backtesting framework (how to verify a strategy)

• Backtest before deployment (don’t “exploit” live until tested)
• Ensure the backtest covers multiple market regimes
• Use a data sample size appropriate to the strategy’s timeframe
• Start with a “black and white” baseline, then:
  • Optimize only after the data foundation is solid
• Use backtesting to stress/prove the strategy doesn’t work:
  • “reverse the hourglass”
  • be extremely tough / rigorous
• Use forward testing / validation after backtest:
  • run on out-of-sample/forward periods to validate deployability

2) Example strategy used to illustrate coded signals

• “Golden Cross” example:
  • 50/200 moving averages
  • Buy when short MA crosses above long MA
  • Sell when it crosses below
• Convert to code:
  • Implement the core signal rule + entry/exit logic
  • Define take profit / stop loss
  • Attach signals to price candles
  • Simulate on historical data (e.g., candle CSV from a provider)

3) Risk metrics: MAE/MFE and win-rate improvement

• Mention of MAE/MFE:
  • MFE (Maximum Favorable Execution) = best unrealized move after entry before exit
• Approach described:
  • Measure average MFE to calibrate TP/strategy
  • If average MFE supports a TP of $3, win rate can approach ~80% (as stated)
• Caution:
  • High win rates may be misleading
  • Claim: “Above 61% mathematically it’s almost impossible to maintain” (as stated)

4) Portfolio construction / diversification argument

• Institutional-style approach:
  • Spread exposure across many assets/strategies, not a single trade
• Retail constraint framing:
  • Retail may succeed, but should avoid institutional-style leverage/risk-taking
• Sector/driver logic:
  • Semiconductors driven by common factors → multiple names can be used while managing single-company idiosyncratic risk

5) Pair trading (named explicitly)

• Pair trading: look for discrepancies between related assets, aiming for a mean-reversion relationship (one laggard outperformer reverting toward benchmark behavior)

---

Key recommendations / cautions

• Don’t trust backtests that are optimized to look great (overfitting / “overfitted robots”).
• Survivorship bias and shiny-object effects are major retail failure modes.
• Automated trading requires correct process + data:
  • “There are no magic formulas for backtesting.”
• Edges degrade over time:
  • Watch for 20–25% degradation to decide whether to retire/replace strategies.
• Risk management is the #1 priority, regardless of account size:
  • Funds can’t afford large drawdowns (risk constraints are survival-critical).
• Be cautious with claims based on “technical analysis levels”:
  • Structures like FVG may be largely random under rigorous testing.
• Bots on retail platforms (e.g., MT5/Telegram) are characterized as “nonsense”:
  • mostly overfitting, no real edge.
• Volatility vs speed warning:
  • Gold described as “fast but not very volatile”
  • Nasdaq described as volatile
• Macro matters, but not as “chart-only”:
  • Macro is framed as probability modeling with accounting/risk models—not just visual patterns.

---

Disclosures / disclaimers

• No explicit “financial advice” or regulatory disclaimer is stated in the provided subtitles.

---

Presenters / sources mentioned

• Charlie Haris (guest; also referenced with variant names such as “Protocol/Protoque/Protosctes”)
• MoneyTalk (podcast host/format; host name not clearly stated in subtitles)

Mentioned entities (not necessarily presenters)

• BNP
• AMP / AMP Global Limited
• Citadel
• Renaissance (Renaissance Technologies)
• LTCM
• Binance
• Hyperliquid
• TradingView