Risk Management for Algorithmic Trading

Two traders can run the same entry and exit rules and end up in completely different places: one compounds steadily for years, the other blows up in a bad month. The difference is rarely the signal. It is risk management, the set of rules that decide how much is at stake on every trade and what happens when things go wrong.

This matters doubly for automated strategies. A discretionary trader might instinctively cut size after a rough week. Your algorithm will keep placing full-size orders through the worst drawdown of its life unless you have written down, in advance, exactly when and how it should not.

Position sizing comes first

Position sizing answers one question: when a trade loses, how much of the account goes with it?

The most durable answer is fixed fractional risk: risk the same small percentage of the account on every trade, typically somewhere around 1 to 2 percent. If your stop loss is 5 percent away from entry and you are risking 1 percent of the account, the position is one fifth of the account. The stop distance sets the position size, not the other way around.

Two properties make this the default choice. It scales down automatically in a losing streak, because each risk unit is a percentage of a now-smaller account. And it makes losing streaks survivable in a way that flat sizing does not: at 1 percent risk per trade, ten straight losses, which any real strategy will eventually produce, costs about 10 percent of the account. Painful, and entirely recoverable.

Stops and exits

A stop loss is where the strategy admits a trade is wrong. For an automated system it is not optional: it is the contract that bounds every position's worst case and the number your position sizing is built on.

Place stops where the trade's idea is invalidated, not at a round number of pain. A mean-reversion entry expecting a snap back within a day or two is invalidated when price keeps stretching; a breakout entry is invalidated back inside the range. Stops set tighter than the market's normal noise get clipped constantly and turn a sound idea into a fee generator. Volatility-scaled stops, for example a multiple of average true range, adapt the distance to how much the instrument actually moves.

Exits deserve equal design attention: profit targets, trailing stops, opposite signals, or time-based exits that close anything living past its expected holding period. Whatever the mix, it must be explicit, because the backtest you ran in the previous lesson is only valid for the exact exit rules it tested.

Strategy-level limits

Trade-level rules are not enough, because strategies fail at the strategy level too. Well-run systems carry circuit breakers above the individual trade:

• A maximum drawdown limit. Decide before going live how deep the equity curve may fall, perhaps one and a half times the worst backtested drawdown, before the strategy switches off for review. Hitting it does not always mean the edge is gone, but it always means the live behavior has left the map the backtest drew.
• Daily or weekly loss caps that pause trading after an unusually bad stretch, which also contains failures that are technical rather than statistical, like a data feed misbehaving.
• Exposure caps on how much of the account may be deployed at once across all open positions.

When Horizon deploys a strategy live, guardrails like stop losses, position limits, and drawdown limits are part of the deployment itself, so the safety rules run with the same discipline as the entries.

Portfolio-level risk

Once several strategies run at once, a new risk appears: correlation. Three strategies that all buy crypto momentum are, in a crash, one large position wearing three names. Diversification only works when the things being combined actually behave differently: different strategy families, different instruments, different timeframes. Combining a trend follower with a mean reverter on different markets genuinely smooths the combined curve; running five flavors of the same idea does not.

Keep a portfolio-level drawdown limit too. The account does not care which strategy the losses came from.

The mindset

Risk management is the part of trading where you are graded on the worst day, not the average one. Entries and exits determine how often you are right; sizing and limits determine whether being wrong is an event or an ending. Get the boring rules right and the strategy earns the time it needs for its edge to play out.

With rules, testing, and risk covered, the remaining step is making it all run without you: How to Automate a Trading Strategy.

Frequently asked questions

How much should I risk per trade?

The standard answer is 1 to 2 percent of the account per trade, defined as the loss if the stop is hit. At 1 percent, a ten-loss streak costs about 10 percent of the account, which is painful and fully recoverable. At 10 percent per trade, the same streak is close to ruin.

Do automated strategies need stop losses?

Yes, more than manual trading does. An algorithm will hold a losing position indefinitely unless a rule closes it, and the stop is also the number position sizing is calculated from. Without one, per-trade risk is undefined.

What is a good maximum drawdown limit?

A common rule is to switch the strategy off for review if live drawdown reaches about one and a half times the worst drawdown in the backtest. Beyond that point the live behavior has left the territory the test mapped, and continuing is no longer evidence-based.