Algorithmic trading risk management

Risk management is the part that has to work

A strategy's edge is uncertain. Its risk controls are not — they are arithmetic, and they either hold or they do not. This is why risk management deserves more care than signal generation, and why it usually gets less.

The purpose of every limit below is survival. A system that survives its bad periods can benefit from its good ones; a system that does not is finished regardless of how good its strategy was.

Size positions from volatility, not from a fixed distance

A stop placed a fixed number of points away means something different in every market and in every week. Twenty points is a wide stop on a quiet currency pair and a rounding error on Bitcoin. A strategy using fixed distances is really running a different strategy in every condition.

Deriving the stop from the market's own recent range — average true range is the usual measure — makes the risk consistent. The stop widens when the market is volatile and tightens when it is calm, so the probability of being stopped by noise stays roughly constant. Position size then follows from the stop distance and the amount you are willing to lose.

The rule of thumb is to risk 1-2% of the account on any single trade. On a small account that produces uncomfortably small numbers, and the discomfort is precisely what empties small accounts.

Set the consecutive-loss limit with arithmetic, not instinct

The probability of a run of k losses is (1 − win rate) to the power k. This is worth actually calculating, because instinct is badly wrong about it.

At a 35% win rate — perfectly healthy for a strategy with a reward-to-risk above 2:1 — four losses in a row happen about 18% of the time. That is roughly one sequence in six. A system that halts on four consecutive losses will halt constantly during normal operation, and the real damage is not the interruption: it is that the operator learns to clear the alarm without reading it.

A limit that means something needs to be rare. At the same win rate, twelve consecutive losses happen about 0.6% of the time. That is a number worth stopping for.

A pause and a kill switch are different tools

This distinction is worth building explicitly. A pause suspends trading for a period and then resumes on its own — the right response to an ordinary run of bad luck, which a strategy winning 40% of the time will have regularly. A kill switch stops the system and waits for a human, which is the right response to something genuinely abnormal.

Put the pause threshold below the kill switch threshold, or the automatic path never runs — the sticky switch fires first and every bad patch needs a person. That inversion is easy to create by accident and produces a system that claims to be automated but needs attention several times a day.

• Pause: after a moderate run of losses inside a time window, or a drawdown over a window. Resumes automatically.
• Kill switch: on a daily loss limit, a balance floor, or a genuinely unusual losing streak. Requires a human.
• Both: logged, with the reason recorded, so the pattern is visible later.

Limits every automated system should have

• A stop loss on every position, placed at the broker rather than held in the bot's memory. A stop that only exists in a running process is an intention, not a stop.
• A maximum daily loss, in money rather than percent, so you know exactly what a bad day costs.
• A cap on trades per day. Frequency grinds an account down through spread as reliably as size does.
• A cap on simultaneous open positions. Two correlated positions is doubled risk wearing a diversification label.
• A balance floor below which the system refuses to trade at all.
• A time-based pause after a bad run that resumes without intervention.

Watch for the strategy degrading

Separate from per-trade risk is the question of whether the strategy still works. Markets change, and a system that was profitable can stop being so without any single trade looking wrong.

A reasonable guard is to track live performance over a rolling window of recent trades and stand down if it falls below a floor. The window has to be long enough that ordinary variance does not trip it — thirty trades is a minimum and more is better — and the floor has to be genuinely poor rather than merely disappointing.

The judgement call is that this cannot be too sensitive, because a strategy whose profit is concentrated in a few large winners will look terrible right before the trade that pays for the month. Standing down on variance removes exactly the trades the edge depends on.

Common questions