The moment that made me stop trusting individual tear sheets was watching ten strategies that each looked fine on their own drop into the same drawdown on the same day. Every one of them had a clean solo backtest. Positive expectancy, reasonable Sharpe, drawdowns I could live with. I stacked them onto one account expecting the losses to cancel out the way the math promised, and instead they all leaned the same direction at the same time and the account took a hit none of the individual sheets had warned me about. Nothing was broken. Each strategy did exactly what it was designed to do. The problem was that I had never tested what they do together.
Testing a strategy in isolation answers a narrow question. Does this rule set make money on its own, with its own capital, with no one else competing for the same fills. That is a fine question to start with, and it is where most people stop. The trouble is that you never actually trade a strategy in isolation. You trade a book. And a book has behaviors that none of its parts have.
What breaks when you put them on one account
There are roughly three ways a collection of individually healthy strategies goes wrong once they share capital, and they are worth naming because you can test for each.
The first is correlated drawdown. Two strategies can look uncorrelated on average and still be tightly correlated in the tail. A momentum system and a breakout system feel like different ideas, but on the day a trend snaps they both bleed at once, because they are both long the same regime. Average correlation hides this. You care about what happens in the worst five percent of days, and that is usually much uglier than the headline number suggests.
The second is margin contention. Each strategy sizes its positions assuming it has the capital available. Run them together and they draw on the same margin pool. When several want to add risk in the same window, you either can not fund all of them or you fund them and quietly run at a leverage none of the solo backtests assumed. The isolated sheet had the whole account to itself. The real book does not.
The third is competition for fills. This one bites hardest in less liquid names and in fast markets. Two of your strategies fire a buy on the same instrument in the same minute. In the solo backtest each one filled at a clean price. Together they are walking the book up against themselves, and the second fill is worse because the first one moved it. You paid slippage to trade against your own account.
Simulate them jointly, on one shared balance
The fix is not complicated to describe, though it takes work to build. You run all the strategies through a single simulation that shares one capital pool, one margin model, and one order book. Not ten separate backtests you average afterward. One engine, one account, all the strategies firing into it on the same clock.
A few things have to be true for this to be worth anything.
- Every strategy runs on the same timeline with the same bar resolution, so their signals land in the right order relative to each other. If one is on daily bars and another on hourly, you have to decide how they interleave before you learn anything.
- Orders compete for the same simulated liquidity. When two strategies hit the same instrument, the second one eats the slippage the first one caused. If your simulator fills both at mid, you have learned nothing about the failure mode that matters most.
- Margin and available cash are tracked at the book level, checked before every entry. A strategy that wants to add risk when the pool is drawn down should get sized down or rejected, exactly as it would live.
- Position sizing reads from portfolio equity, not a fixed notional per strategy. Otherwise you are not testing allocation, you are testing ten islands that happen to share a spreadsheet.
Once you have that, the output is different in kind from a stack of tear sheets. You get one equity curve for the book, and you can start asking the questions that actually decide whether you should run these together.
Measure what each strategy adds to the risk, not to the return
Marginal contribution is the idea I lean on most here. The question is not how much a strategy makes on its own. It is how much risk the book takes on because that strategy is in it. Two strategies can have identical solo returns while one of them roughly doubles your worst-week drawdown and the other quietly dampens it.
The way to see this is to run the joint simulation with the full set, then pull one strategy out and run it again. Watch what happens to the portfolio numbers you care about, which for me are max drawdown, the depth of the worst rolling week or month, and how long the book stays underwater. A strategy that reduces those when you add it is earning its slot even if its standalone return is modest. A strategy that inflates them is expensive no matter how good its solo sheet looks. Do this leave-one-out pass for every strategy and you get a ranking that has almost nothing to do with individual returns.
The other thing worth watching is the crowding moment. Log every time two or more strategies want the same instrument in the same direction within a short window. Those clusters are where correlated drawdown and fill competition both live. If you find that half your bad days trace back to three strategies piling into the same trade, you do not have five ideas, you have one idea wearing five costumes, and you are paying five times the slippage to express it.
A rough workflow for allocating across the book
Here is roughly how I go from a pile of passing strategies to an allocation I trust.
- Run every candidate through the joint engine on shared capital, same clock, competing fills. Throw out anything whose numbers fall apart the moment it shares a book.
- Do the leave-one-out pass. Rank strategies by marginal contribution to drawdown, not by solo return. Flag any that make the book meaningfully worse.
- Look at the crowding log. Where several strategies keep wanting the same trade, either cut the redundant ones or cap how much of the book any single instrument can absorb.
- Size allocation by the joint behavior. Give more room to the strategies that dampen the book and less to the ones that amplify it, even when that means underweighting your best solo performer.
- Stress the whole book together. Push it through the ugliest historical windows you have and a few synthetic ones where correlations go to one, which is what they do in a real panic.
The uncomfortable part is that this process regularly tells you to cut a strategy you like. Something with a beautiful standalone curve turns out to add risk the book does not need, and the honest move is to bench it. I have done that more than once and it still stings.
None of this replaces testing strategies in isolation. You still need the solo pass to know a strategy has any edge at all. It just stops being the last word. A tear sheet tells you whether an idea works alone. The joint simulation tells you whether it belongs in the room with the others, and that is the question that decides how the account actually behaves. Start with the solo test to earn a seat, then let the portfolio simulation decide how big the seat gets.