Risk management

A strategy with positive expected return and no risk control is a slot machine with one extra step. Risk management is the discipline of deciding, before you take a position, how large it should be, how much loss you will tolerate before you cut it, how the position relates to everything else you hold, how much leverage you can survive, and at what point you stop trading entirely. None of these decisions are about predicting the market. All of them are about surviving long enough for the predictions to play out.

Position sizing

The first question is not what to trade but how much. Three frameworks dominate practice.

Kelly criterion. Kelly (1956), working at Bell Labs on a signal-noise problem for telephone lines, derived the bet size that maximises the long-run geometric growth rate of capital. For a binary bet with win probability p, win size b, and loss size a, the Kelly fraction is (p b - (1-p) a) / (a b). In trading the analogue is the fraction of capital allocated to a position; for a strategy with expected excess return μ and variance σ² the continuous-time approximation is f* = μ / σ². The intuition that matters is that Kelly bets large when edge is large and variance is small, that betting more than Kelly is suicidal in the long run (the geometric growth rate goes negative for f > 2f*), and that betting less than Kelly — typically half-Kelly or quarter-Kelly — gives up some growth in exchange for much lower volatility. Almost no working quant bets full Kelly; the variance in estimated edge alone justifies a large haircut.

Fixed-fractional. Allocate a constant fraction (commonly 1% to 5%) of current capital to each position. Simple, scale-invariant, and good enough for many discretionary and systematic strategies. The downside is that it ignores the volatility of the underlying — a 2% allocation to a sleepy utility and a 2% allocation to a meme stock are wildly different exposures.

Volatility targeting. Size the position so that its contribution to portfolio volatility is constant. If you target 10% annualised portfolio volatility and the position has 40% annualised volatility, the position weight is 10% / 40% = 25%. This is the modern institutional default because it adapts automatically to changing market regimes: when realised volatility doubles, position size halves.

AlphaHub action: every backtest report shows the realised annualised volatility of the strategy. Use it to back into the implicit Kelly fraction (μ / σ²) and to check whether you would actually accept that bet size on capital you cared about.

Drawdown limits

A drawdown is the percentage loss from the most recent equity peak. The maximum drawdown over a backtest is the worst day you would have lived through if you had been trading the strategy in real time. Most discretionary investors capitulate at 15-20% drawdown; institutional allocators typically redeem at 25-30%. A strategy with a great Sharpe and a 50% maximum drawdown is in practice unrunnable because the human in front of the screen will fire it before it recovers.

The headline number to watch alongside Sharpe is the Calmar ratio: annual return divided by absolute maximum drawdown. A strategy with 15% annual return and 30% max drawdown has a Calmar of 0.5; one with 8% annual return and 8% max drawdown has a Calmar of 1.0 — the second is more pleasant to live with even though the first earns more. Hedge-fund allocators commonly require Calmar > 0.5 for serious capital and Calmar > 1.0 for "comfortable" allocations.

Soft stops scale down exposure as drawdown grows (cut to half-size at -10%, quarter-size at -15%). Hard stops flatten the book entirely at some threshold (-20%, say) and require a defined re-entry condition (drawdown back inside -10%, three positive months, etc.) rather than discretion. Either works; what fails is having no rule at all.

AlphaHub action: the backtest metrics include max_drawdown_pct and sharpe. Compute Calmar yourself as ann_return / abs(max_drawdown_pct) and require it to be at least 0.5 before taking the strategy seriously.

Correlation

Markowitz (1952) showed that the variance of a portfolio depends as much on the correlations between assets as on the variances of the individual assets. The practical consequence: ten "different" long positions that all turn out to be levered exposures to the same risk factor are equivalent to one big position, and the diversification you thought you had does not exist. The 2008 mortgage crisis was a giant lesson in this — AAA-rated tranches across hundreds of mortgage pools all defaulted together because the pools shared a single underlying factor (US housing).

Working diagnostics: compute the pairwise correlation matrix of position returns during the backtest, look at the largest eigenvalue (it captures the dominant factor your portfolio is loaded on), and run a stress test where every correlation goes to 1 (the "correlation breakdown" scenario, which is exactly what happens in liquidity crises). A portfolio that survives the all-correlations-to-1 stress at acceptable drawdown is genuinely diversified; one that does not is concentrated in disguise.

AlphaHub action: when you backtest a multi-name strategy, the report's per-symbol contribution table shows which names drove the cumulative PnL. If two or three names dominate, your strategy is more concentrated than the name count suggests; size accordingly.

Leverage

Leverage amplifies both return and risk, but not symmetrically. The asymmetry is the volatility drag of compounding: the long-run geometric return of a leveraged strategy is approximately L μ - L² σ² / 2, where L is leverage, μ is the unleveraged expected return and σ is the unleveraged volatility. The drag term grows with the square of leverage, so 2x leverage doubles return but multiplies drag by 4. For a strategy with μ = 8% and σ = 20%, the optimal leverage for geometric growth is μ / σ² = 0.08 / 0.04 = 2.0; pushing past 2x reduces compound growth even though arithmetic return is still increasing.

The drawdown amplification is steeper. A strategy with a 20% unleveraged max drawdown levered 2x typically reports closer to a 45% max drawdown (not 40%) because drawdown clusters in periods of elevated volatility, and the leverage interacts with the volatility cluster. The empirical "drawdowns scale roughly as L^1.5" is a useful rule of thumb when sizing leverage for a target risk budget.

AlphaHub action: AlphaHub's backtest applies leverage as a target_weight multiplier; if the unleveraged report shows max_drawdown = 18%, expect the 2x version to show roughly 30-40% — verify before deciding.

Kill switches and cool-down rules

Even good strategies have bad streaks. The question is whether a bad streak is a draw from the strategy's normal distribution (live with it) or evidence that the strategy has structurally broken (stop trading it). Distinguishing the two in real time is hard; the safest default is to define both an exposure floor and a cool-down rule in advance.

Examples of rules that working desks use:

• Trailing-loss kill: if the drawdown from peak exceeds 1.5x the worst drawdown observed in the backtest, flatten the book and pause for review. The 1.5x multiplier reflects that out-of-sample drawdowns are systematically worse than in-sample ones.
• Losing-streak cool-down: after N consecutive losing days (calibrated to a 1% tail of the backtest's losing-streak distribution), reduce exposure by half until the next winning day.
• Volatility ceiling: if realised portfolio volatility exceeds 2x the target, scale all positions down until it returns. This catches the case where the model's variance estimate has gone stale.

Lopez de Prado (2018), in chapter 15, argues that bet-sizing rules and stop-loss logic should be designed jointly with the alpha signal rather than bolted on afterwards; in particular, that "stop-loss" levels should be set as a function of the signal half-life (how long the alpha persists) and the prevailing volatility, not as a flat percentage of capital. The detail matters less than the principle: rules written down in advance survive panic better than judgement made under it.

Take my last five backtests. For each, compute the Calmar ratio (annual_return / max_drawdown_pct) and the implied half-Kelly fraction (0.5 * sharpe / annualised_volatility). Rank them by Calmar and tell me which is the most allocator-friendly.

References

• Kelly, J. (1956). A new interpretation of information rate. Bell System Technical Journal, 35(4), 917–926.
• Lopez de Prado, M. (2018). Advances in Financial Machine Learning. Wiley. Chapter 15 on bet sizing and stop-loss design.
• Markowitz, H. (1952). Portfolio selection. Journal of Finance, 7(1), 77–91.