Discussion of drawdown is descriptive of how the system was designed and how it has performed historically. It is not a guarantee of any particular forward drawdown, and any live-trading account will experience drawdowns of unknown magnitude. Past performance does not predict future results.
Most quant marketing leads with return. The serious money leads with drawdown. The return number you see is the sample; the drawdown number is the contract.
Why the Sharpe ratio under-states the problem
Return divided by volatility is a clean academic construction and a deeply unhelpful operational one. Two strategies with the same Sharpe ratio can have radically different equity curves: one that wanders into a 50% drawdown for thirteen months and one that never crosses 15%. The Sharpe doesn't distinguish them; the institutional allocator does. The institutional allocator never gets the chance to live through the first one, because the redemption notice arrives at month six.
Drawdown is what kills strategies in production. Not bad returns, not high vol — drawdown. The size, depth, and duration of the worst peak-to-trough is what determines whether a system survives the operational reality of being run. It is also, conveniently, the one risk metric that compounds with the return: a 50% drawdown requires a 100% recovery to break even. The asymmetry is exactly the opposite of the one favoured by reflexive Sharpe maximisers.
How Shishin treats drawdown
Drawdown is a first-class metric in the Shishin design loop. Every parameter change, every weight adjustment, every proposed new engine is evaluated on maximum drawdown alongside CAGR. A change that improves CAGR by 200 basis points but worsens drawdown by 500 is rejected. A change that flattens the equity curve materially without giving up more than a few hundred basis points of return is shipped. The trade is asymmetric in the system's favour: we will give up a real amount of headline return to make the curve wearable.
Three mechanisms produce most of the drawdown compression:
- ATR-based exits. Every entry ships with a stop derived from the recent volatility of the underlying, not a fixed percentage. A name with a 6% ADR gets a wider stop than a name with a 2% ADR — same dollar risk per name, different distance. The stop is set once and not moved adversely; only trailed when the trade is in profit.
- Position cap + score-gated bypass. The live stack never holds more than a small fixed number of concurrent positions. The cap can be exceeded only by names with composite scores in the top decile of the live universe, which forces capital concentration on high-conviction names rather than diluting it across a long roster of mediocre ones.
- Engine rotation by macro. The single most important mechanism. When the broad market deteriorates, the macro switch rotates out of the small-cap momentum engine (Suzaku, peak standalone drawdown 32.5%) and into either defensive-sector vehicles (Byakko, peak 33.1%) or recovery large-caps (Seiryū, peak 40.9%) or simply to cash. Each engine on its own can produce a drawdown north of 30%. The live stack's worst was 17.7%.
The gap is the architecture
The interesting number isn't the live stack drawdown by itself. It's the gap between the live stack drawdown and the worst per-engine drawdown. Suzaku alone touched 32.5% in the standalone backtest. The live stack — the same Suzaku engine, plus the other three, gated by the macro switch — topped out at 17.7%. The delta between those numbers is what the four-engine architecture buys. It is also the answer to the most common question about regime-switching systems: is the macro switch worth the added complexity? The drawdown gap is the worth.
What we don't do
We don't use stop-losses tighter than the recent volatility of the underlying. Tight stops generate high noise hit rates — the position is exited on a normal price oscillation rather than a thesis break — which front-loads losses without actually reducing tail risk. We don't average down. We don't re-enter a stopped-out trade within a holding-period window. We don't run portfolio-level kill switches that flatten everything on a single bad session, because the operational cost of an unnecessary liquidation (slippage, lost compounding, the noise it injects into the live track) is higher than the cumulative tail risk it averts.
And we do not optimise for return. We optimise for the shape of the curve, with return as a constraint. A strategy that does not respect the shape of the curve does not survive the conditions under which it is asked to make money. Drawdown is the metric the system pays its respects to. The CAGR follows from doing that well, not from targeting it directly.