Table of Contents
Ernest Chan—quant researcher, hedge fund manager, and author—sits down for a practical talk on building systematic trading the right way. From launching QTS Capital to engineering his intraday “Tail Reaper” approach, Ernest explains how a physicist’s mindset meets real-world execution, why overfitting wrecks naive AI, and how machine learning can enhance (not replace) a trader’s judgment.
In this piece, you’ll learn how Ernest designs a strategy from idea to live trading—reading research, sanity-checking edge after costs, paper-trading, then scaling with tight risk controls. You’ll see how he uses machine learning to filter low-quality trades, select meaningful features, and hedge tail risk without relying on overnight exposure, plus his guardrails for avoiding the “black box” trap and staying adaptive through drawdowns.
Ernest Chan Playbook & Strategy: How He Actually Trades
Markets & Instruments He Chooses
Ernest Chan keeps it simple: liquid, diversified, and cheap to trade. The goal is to reduce frictions so the edge—not fees—drives P&L. Think major futures and highly liquid equities where execution quality is reliable.
- Focus on liquid futures (ES, NQ, CL, GC, ZN, 6E) and top-1000 U.S. equities by dollar volume.
- Require average bid-ask spread ≤ 1 tick and depth ≥ 3x your typical order size.
- Hard-exclude single names around earnings; flatten 24 hours before and after.
- Assume conservative costs in research: 0.5–1.5 ticks per round trip in futures; 2–5 bps per side in equities.
- Trade intraday where feasible; avoid overnight gaps unless the strategy is explicitly designed for it.
How He Finds Edge (Mean Reversion & Momentum)
He hunts for simple, testable patterns first, then pressures them with costs and robustness checks. Mean reversion dominates in equities; momentum and breakouts carry more weight in futures.
- Equities mean reversion: buy when close < k-sigma below a short moving average (e.g., 5–10 bars); sell when it snaps back to the mean.
- Futures momentum: enter on break of prior day high/low or moving average cross with volatility filter (e.g., ATR(14) > 20-day median ATR).
- Only keep signals that survive slippage assumptions and produce positive net expectancy out of sample.
- Use simple features first (returns, ATR, range, session time); add complexity only if it adds net edge.
- Cap simultaneous correlated bets: max 3 names per highly correlated cluster (e.g., semis, homebuilders).
Machine Learning Without the Magic
For Ernest, ML is a filter and ranker—not a crystal ball. He uses it to prioritize trades and reduce noise, but he insists on rigorous validation to avoid overfitting.
- Treat ML as a ranking model: score setups and execute only top decile/quintile.
- Use time-series CV (purged, embargoed splits) to stop label leakage; never random shuffle.
- Limit features to stable, economically sensible ones; drop anything with regime-dependent drift.
- Regularize aggressively (L1/L2/elastic net); prefer tree ensembles with depth limits over deep nets.
- If live hit-rate drops > 10 percentage points from the CV average for 60 trading days, disable the ML filter and revert to base rules.
Risk & Position Sizing
Sizing is the throttle. Ernest leans on volatility targeting and a conservative fraction of Kelly to keep drawdowns tolerable and compounding steadily.
- Volatility-target each position to a fixed daily risk (e.g., 10–25 bps of equity per trade).
- Position size = target_risk / recent_vol; use 20-day realized or intraday ATR for vol.
- Portfolio risk cap: 1.0–1.5% of equity at 95% one-day VaR; never exceed 2% under any scenario.
- Use 0.25–0.5× Kelly on proven signals; cut to 0.1× Kelly during volatility regime shifts.
- Daily stop: close book if combined open losses hit 0.75–1.0% of equity; no “revenge” trades.
Execution: Turn Edge Into Fills
He treats execution as a strategy layer. Reduce slippage and rejection rates, and you keep more of the modeled edge.
- Use limit-at-touch with protective market-if-touched when liquidity is deep; flip to marketable limits during news-spike exits.
- Slice orders: child size ≤ 10–20% of displayed top-of-book to avoid footprinting.
- For equities, avoid opening 5 minutes and closing 10 minutes unless the model is built for it.
- Enforce a max slippage per fill (e.g., 0.3–0.7 tick futures; 1–2 bps equities); cancel/repost if breached.
- Flatten intraday strategies by session end; if holding overnight, hedge beta/delta to target exposure.
Portfolio Construction & Diversification
Single strategies come and go; portfolios compound. Ernest spreads risk by instrument, timeframe, and logic (MR vs. momentum vs. carry/tail).
- Combine at least two orthogonal edges: equities mean reversion + futures momentum as a baseline.
- Target cross-strategy correlation < 0.5; drop or down-weight if correlation > 0.7 over 60 days.
- Equal risk weight by strategy sleeve (risk parity across sleeves), not equal capital.
- Cap any sleeve at 40% of total risk; cap any single symbol at 10% of total risk.
- Re-optimize sleeve weights monthly using realized volatility and drawdown penalties.
Tail Risk: The “Tail Reaper” Mindset
When markets snap, he wants to be long convexity, not excuses. The goal is to monetize large intraday trends and volatility expansions, then get flat.
- Activate breakout models when realized vol rank > 70th percentile or VIX crosses a threshold you predefine.
- Place stop-entry orders ±k×ATR from reference price; trail exits at 1.5–2.5×ATR or time-based EOD.
- Restrict countertrend trades on high-vol days; only follow the direction of the initial expansion.
- Hard daily give-back rule: if you return > 35–50% of peak open profit, exit the remainder.
- Keep tail sleeve independent: it runs when vol explodes; turn it off when vol normalizes.
Research → Production Pipeline
He pushes ideas through a strict funnel so only robust, capacity-aware strategies make it live. Paper profits don’t count; live stats do.
- Prototype with conservative costs; reject if in-sample Sharpe < 1.0 before costs or < 0.6 after costs.
- Demand out-of-sample and forward-walk Sharpe within 30% of in-sample; otherwise, discard.
- Paper trade for 20–40 sessions; promote to small capital if slippage and rejects match assumptions.
- Graduated scaling: 1× → 2× → 3× size only after each step clears slippage and drawdown gates.
- Version control model: only one live change per strategy per week to avoid thrash.
Regime Detection & Switching
Markets change character. Ernest tracks regimes and adapts models and risk, so he’s not forcing yesterday’s rules onto today’s tape.
- Define regimes using volatility state, term structure (e.g., VIX/VIX3M), and trend filters on risk assets.
- Maintain per-regime parameters (lookbacks, ATR multipliers, time-of-day behavior).
- Auto-deactivate mean reversion in crash regimes; elevate momentum/tail sleeves.
- Require a minimum dwell time (e.g., 10–15 sessions) before switching back to prior parameters.
- Log regime transitions and P&L attribution to confirm the switch added value.
Live Monitoring & Kill-Switches
He treats risk like a system, with pre-committed thresholds that force action before emotions take over.
- Track rolling 60-day live Sharpe, max drawdown, and win-rate; disable any strategy if two of three breach limits.
- Kill-switch examples: 60-day Sharpe < 0, win-rate < base by 10 p.p., drawdown > tested by 1.25×.
- Enforce symbol-level circuit breakers after repeated adverse slippage or news events.
- If daily P&L < −1.0% equity, shut down discretionary overrides; only allow protective exits.
- After a kill, require a structured post-mortem and a 10-day paper period before reactivation.
Practical Settings & Defaults (A Good Starting Grid)
Defaults aren’t laws; they’re safe places to begin testing. Ernest favors conservative assumptions that survive real execution.
- Lookbacks: MR (5–10 bars intraday; 2–3 days swing), MOM (20/50-bar cross or prior-day extremes).
- Vol filters: trade only when ATR(14) > 20-day median ATR; widen stops/targets as ATR rises.
- Targets/stops: 1.0–1.5×ATR target for MR; 2.0–3.0×ATR trail for momentum/tail.
- Sizing: 10–25 bps of equity risk per position; 0.25–0.5× Kelly on validated edges.
- Review cadence: weekly parameter sanity check; monthly sleeve weights; quarterly full-book audit.
Size trades by volatility; throttle risk with conservative Kelly fractions
Volatility is the speed limit, and Ernest Chan treats it like law. He sizes positions so a fixed fraction of equity is at risk per trade, using recent realized volatility or ATR to translate price noise into dollars. When markets heat up, size ratchets down; when they cool, size can breathe a bit. The result is steadier equity-curve behavior and fewer wipeouts from “normal” volatility spikes.
Ernest Chan also resists the siren song of full Kelly, preferring a conservative fraction to keep drawdowns livable. He’ll estimate edge and variance, then run at roughly a quarter to a half of Kelly, so bad streaks don’t derail compounding. If live results deviate from expectations, he cuts size first and asks questions second. This discipline makes risk a knob he can turn—not a surprise that turns him.
Build edges with simple rules, validate brutally, and deploy only robust.
Ernest Chan starts with rules you can explain on a whiteboard: mean reversion triggers, breakout levels, and volatility gates. The simplicity is deliberate—clean logic is easier to test, maintain, and fix when markets change. He pressures every idea with realistic costs and slippage so paper edges don’t vanish in the real world. If a rule needs heroic parameter tuning to “work,” Ernest bins it.
Validation is where most strategies die for Ernest Chan, and that’s the point. He demands out-of-sample performance within striking distance of in-sample, plus forward tests that match the backtest’s character after costs. Regime shifts are simulated with stress windows and walk-forward splits to make sure the rule doesn’t rely on one market mood. Only after the paper fills line up with assumptions does he allocate small capital and watch live stats like a hawk. If the live hit rate or Sharpe ratios are beyond pre-set bands, the strategy is cut or fixed before it can bleed the book.
Diversify by instrument, timeframe, and logic; cap correlation across sleeves.
Ernest Chan spreads risk across what truly matters: different markets, different holding periods, and different ways of extracting edge. He’ll pair equity mean reversion with futures momentum, so the book isn’t leaning on one market mood. Timeframes are staggered—intraday, swing, and occasional multi-day holds—so the same shock doesn’t hit every position the same way. Ernest Chan watches cross-sleeve correlation like a hawk; when it climbs, he trims or reweights before drawdowns sync up.
Diversification isn’t a sticker—it’s an active process in Ernest Chan’s playbook. He caps any single sleeve’s risk share and limits symbol concentration, especially inside highly correlated sectors. When volatility clusters, he reduces exposure where correlations spike and lets uncorrelated sleeves carry the load. The goal is a portfolio that composes edges like a band, not a solo act—if one instrument or logic goes off-key, the rest can still keep the rhythm.
Let momentum run, take mean reversion snaps; avoid overnight gap risk.k
Ernest Chan treats momentum and mean reversion as different gears in the same car. When price expands with rising volatility and broad confirmation, he lets winners breathe instead of grabbing quick profits. Mean reversion gets a faster trigger and a tighter leash—enter on stretched moves, exit at the mean, and don’t overstay. Ernest Chan emphasizes that the edge comes from matching hold time to the behavior: momentum wants trails and time, reversion wants precise entries and swift exits.
Overnight is where random gaps can nuke carefully modeled risk, so Ernest Chan favors intraday bookkeeping unless the strategy is explicitly gap-aware. He’ll flatten momentum trades by the session’s end if the thesis was intraday and reprice in the morning rather than gamble on news. For mean reversion, he avoids holding through earnings or macro events that turn small edges into coin flips. The combination is simple but strict: press trends while they move, snap back when they stretch, and refuse exposure to gap risk that your rules didn’t price.
Use machine learning to rank setups, not predict prices; kill quick.ly
Ernest Chan treats machine learning like a bouncer at the door, not the DJ. The job is to score and rank otherwise valid setups, letting only the highest-quality trades in while the rest wait outside. He sticks to sensible, stable features and demands that the top-ranked bucket materially outperforms the median in forward tests. The payoff is fewer low-quality entries and a tighter hit-rate, without pretending the model can foretell every tick.
When the model drifts, Ernest Chan pulls the plug fast. He watches live bucket spreads, hit-rate, and slippage against pre-set tolerances; if two metrics break, the ML filter is disabled and base rules take over. Re-training happens on a schedule with embargoed splits, and new versions must beat the old one net of costs before seeing live capital. The mantra is simple: use ML to prioritize execution, monitor it like a hawk, and cut it the moment it stops adding value.
Ernest Chan’s key lesson is that robust trading sits on a scientific process and strict risk control. He learned the hard way that leverage can turn even strong edges into painful drawdowns, which pushed him to engineer intraday, no-overnight strategies that can hedge tail events while avoiding gap risk—most notably his “Tail Reaper” approach that trends with volatility spikes and stands alone as a crisis sleeve. From there, he refines edges with machine learning as a filter, not a fortune teller, using it to limit losses and even eke out small gains during calm markets rather than trying to predict every tick.
Another core takeaway is the “human first, ML second” workflow. Ernest argues that markets don’t yield clean labels, so a trader must craft the base strategy—define the signal, the logic, the risk—before any algorithm starts ranking opportunities; the machine should learn from the trader’s framework, not “from the market” directly. That framework is built in a lab-style loop: read widely, generate hypotheses, punish them with realistic costs and liquidity limits, and only advance ideas that survive out-of-sample and forward tests—because most published edges don’t.
Finally, he’s explicit about where he’s pushing next: volatility trading with futures and options, paired with continued commercialization of his ML tooling—an evolution consistent with his broader message that diversification by logic and regime is what keeps a book alive. Underneath it all is the mindset of a physicist who treats markets as the toughest pattern-recognition problem on earth: stay curious, size by volatility, prefer intraday when your rules demand it, and let ML help you avoid bad trades rather than blind you to risk.
Related posts:
- Elliot Surge: Trader Strategy Without The Hype
- Trader Mindset Strategy: How Two Dentists Built a Billion-Pound Machine
- Trader Strategy Spotlight: Albert Burgess on Process Over Hype
- Pat Frain Trader Strategy: How a 21-Year-Old Built Consistency Over Hype
- Rizz the Trader Strategy Interview: How Real Value Beats Hype
- Adam Grimes Trader Strategy: Data-Driven Discipline Without the Hype
- Data-Driven Trader Strategy: From Crypto Hype to Real-Edge Execution
- Trader Strategy Showdown: Prop-Firm Hype vs Real Desk Reality
- FTMO Record Holder’s Trader Strategy: Bernd Skorupinski on Real Edge vs. Prop Hype
- Trader Strategy Spotlight: Trader Nick on Building Real, Steady Edges
- Adam Scarr Trader Strategy: How a Funded SMC Trader Built Confidence and Consistency
- Jerremy Newsome Trader Strategy: How a Pro Day Trader Builds Rules, Mindset, and Momentum
