If you read one page on this site, read this one. It's not rhetoric: my public real account crossed a 72% drawdown, and that number — not the +458% next to it — is why every one of my systems now has risk constraints written into the code. Risk management isn't the boring chapter after the strategies: it's the reason you'll still be trading in three years, or not.
The asymmetric math of losses
| Loss | Gain needed to break even |
|---|---|
| −10% | +11% |
| −25% | +33% |
| −50% | +100% |
| −72% | +257% |
| −90% | +900% |
Beyond a certain threshold, recovery is mathematically possible but statistically unrealistic — and psychologically devastating. Hence the first law: drawdown is sized beforehand, not commented on afterwards.
How much to risk per trade
The professional standard is 0.25%–1% of capital per trade. The reason is arithmetic, not prudence: losing streaks are a normal event in any system (a 50% win-rate system has over 99% probability of hitting 7 straight losses within 1,000 trades). At 1% per trade, 7 straight losses = −6.8%: annoying, manageable. At 5% = −30%: one step from the point of no return, reached by a perfectly ordinary sequence.
The position sizing formula
Size = (Capital × Risk%) / (Stop distance × Point value)
Example — $10,000 account, 1% risk (= $100):
50-pip stop on EURUSD, 1 mini-lot pip = $1
Size = 100 / (50 × 1) = 2 mini lots (0.20)- Stop first, size second. The stop goes where the trade's hypothesis is invalidated (technical level, ATR multiple) — never at a distance chosen to "fit" the size you want. Size is the dependent variable, always.
- Stops in volatility units. A k×ATR stop adapts to the market; a fixed-point stop becomes wrong as soon as volatility changes. The same lesson that broke one of my systems when gold tripled.
- Automate the calculation. Doing it by hand under pressure produces errors — and the prop account burned by a badly rounded lot is a classic. It's exactly why I built FastAutolot.
Kelly, and why it's used fractionally
The Kelly criterion computes the capital fraction that maximizes geometric growth given your edge. The problem: it assumes you know your true win rate and payoff — you only know their historical estimates. Full Kelly on estimated parameters produces brutal swings; the quant standard is fractional Kelly (¼–½), sacrificing theoretical growth for real robustness. If your computed Kelly is above 2–3% per trade, you almost certainly don't have a monstrous edge: you have an optimistic estimate.
The three lines of defense
- Per-trade stop loss: the unit risk, defined above.
- Daily loss limit: when the day hits −2/−3%, close everything and stop. It makes revenge trading mathematically impossible — an architecture problem, not a character problem.
- Max drawdown limit with a latch: if the account hits −X% from its peak, the system stops and stays stopped until a cold-headed review decides. In my EAs it's a code module with priority over everything: it checks equity (not balance), closes all positions and doesn't re-arm itself.
The risk you don't see: the tail
The most dangerous profile isn't the frequent loser: it's the one who almost always wins a little and rarely loses enormously — grids, martingales, averaging down. 80–90% win rate, a smoothly climbing equity, and a hidden tail that eventually collects the account. I keep an entire public demo account to prove it: +641% and mathematically doomed. If your method involves "averaging" losing positions, you don't have risk management: you have a tail generator.
The next level — headroom-anchored sizing, prop constraints, circuit breakers in code — is in the prop firm playbook and the book.