Running More Than One Strategy at Once

The obvious next move after validating one strategy is to run a second one alongside it, on the reasoning that diversification is free improvement. Sometimes it is. Often the two rules are the same bet wearing different code, and combining them concentrates risk while the per-strategy reports both continue to look reassuring — because neither report can see the other.

Educational material, not trading advice. Algorithmic crypto trading is high-risk and most retail algo traders lose money.

Correlate the returns, not the descriptions

Two strategies are distinct if their net return series are weakly related, and for no other reason. Different indicators, bar intervals, code, authors: none of it is evidence. A breakout rule on one interval and a trend filter on another can produce positions that move together almost perfectly, because both are long when the market has been going up.

So the first thing to compute for any candidate pair is the correlation of their per-period net return series — after costs, same universe, same window. Use returns rather than positions: positions of different sizes contribute unequally, and returns are what add up.

Two refinements. Look at correlation in the losing periods separately, because strategies frequently decouple in calm conditions and converge in bad ones — the opposite of the arrangement you wanted. And check that the correlation is stable across the sample rather than an average of one high-correlation stretch and one low one; an average correlation of a relationship that varies describes nothing.

What diversification does, and doesn’t

Combining reduces the variance of the combination, and does nothing about a shock common to both. The arithmetic is worth having in front of you. For illustration, take two return streams with identical volatility σ and correlation ρ, combined at equal weights. The combined volatility is

σ_combined = σ × sqrt((1 + ρ) / 2)

That formula is the entire argument for and against multi-strategy trading. At ρ = 1 the combination has exactly the volatility of either one — complexity added, nothing gained. At ρ = 0 you get a meaningful reduction. Below zero you get more, at the cost of the rules partly cancelling each other’s returns as well as their risk. Those are hypothetical inputs illustrating the shape of a relationship, not a claim about any real pair of strategies.

What the formula hides is that ρ is not a constant. It is an estimate from a sample, with its own error bar, and correlations tend to rise when conditions deteriorate. Planning capital around a correlation measured in ordinary times means the diversification is present when you don’t need it and absent when you do.

Overlapping exposure is a separate problem

Correlation is about return streams; overlap is about holding the same thing at the same time, and it is a distinct risk that per-strategy sizing cannot see.

If three strategies each size themselves to risk a modest fraction of capital, and all three are long the same instrument on the same afternoon, the account’s exposure to it is three times what any single strategy believes. Each strategy’s risk report is correct and the portfolio’s is not. The failure is structural: sizing decided locally cannot enforce a global constraint.

The fix is a limit that lives above the strategies rather than inside them — a portfolio-level cap on gross exposure, on net exposure, and on exposure to any single instrument, applied to the combined intended position before any order goes out. A strategy asking for more than the cap allows gets scaled down, and which one gets scaled is a policy decided in advance rather than an emergent property of whichever ran first. This is the same argument for limits outside the strategy made in what sits between a backtest and a live bot, and the reason risk control outranks signal quality in position sizing and risk management basics.

Netting is not automatically a win either. Two strategies with opposing positions in the same instrument may net to nothing — which saves the exposure and also means you paid two sets of costs to hold nothing.

Allocating capital between them

Every allocation scheme is a model, and the more it learns from your backtests the more of your overfitting it inherits. Roughly in order of how much they assume:

Equal weight. Assumes nothing except that you cannot rank them, which is more often true than people like. The honest default, and a surprisingly hard baseline to beat out of sample.

Inverse volatility. Weight each strategy by the reciprocal of its realized volatility, so each contributes similar risk rather than similar capital. It needs one estimated number per strategy, and volatility is among the more stable things you can estimate — a reasonable step up.

Optimized weights. Solving for the weights that would have maximized some backtest metric is another parameter search on the same data, with all the properties in multiple testing: why your best result is probably noise. It also needs a full correlation matrix, and correlation estimates are far noisier than volatility estimates, so the optimizer’s confident output rests on the least reliable inputs available. If you do it, do it inside walk-forward so weights are chosen on data preceding the period they apply to — and expect something close to equal weight once the noise is handled.

Backtest them jointly

A portfolio has to be tested as a portfolio, because several important effects exist only in the combination. Separate backtests summed together silently assume infinite capital, independent costs, and no interaction.

What a joint backtest represents that summing does not: shared capital, so simultaneous signals compete rather than both being funded; netted positions, so offsetting trades are not double-charged; portfolio-level limits actually binding and truncating positions; and shared capacity, since two strategies wanting the same exposure at once compete for the same liquidity, so each one’s measured capacity overstates what the pair can hold — see backtesting a strategy that can’t fill at once.

The joint result will be worse than the sum of the individual ones. That gap is not an engine bug; it is the cost of the strategies sharing an account, and you want the number before production hands it to you.

Turning one off

Define per-strategy retirement criteria before adding the strategy, not after it disappoints. A portfolio makes this harder, because a bad stretch in one component is masked by the others — the account looks fine while one rule quietly stops working. So each component needs its own monitored performance and its own pre-committed threshold, judged on its own returns rather than the portfolio’s. Telling genuine decay from an ordinary drawdown is the subject of regime change and why strategies decay; the addition here is that in a multi-strategy account you must look for it deliberately, because the aggregate will not tell you.

The takeaway

Adding a second strategy is not free diversification — it is a new hypothesis about the relationship between two return streams, and that relationship needs measuring, limiting, and testing jointly like anything else. Two questions before every addition: does this behave differently from what I already run, measured on returns and not on intentions? And can my risk limits see the combined position rather than each piece separately? If either answer is no, one more strategy is one more way to lose the same money twice.