Run Several Strategies, Not One: Diversification That Really Cuts Drawdown

Even a good trading system has bad years. The way professional system traders make those bad stretches shallower is not by finding a system that never loses. It is by running several systems that lose at different times. This article simulates how much that helps, which formulas measure it, and why most of the “diversification” you see is fake.

If correlation is new to you (ρ, a number from −1 to +1 that tells how much two things move together), read Correlation and portfolio risk first. That article is about holding several currency pairs. This one is about holding several systems.

Why two ordinary systems beat one equally good system

Say systems A and B have the same average profit and the same daily swing (σ, the standard deviation of daily profit and loss). We split the capital half and half.

• Combined average profit = ½ × A’s profit + ½ × B’s profit = unchanged
• Combined swing, from the formula in the correlation article:

σ_combined = σ × √((1 + ρ) ÷ 2)

Plug in some values:

1. ρ = 0.9: √(1.9 ÷ 2) = √0.95 = 0.975, barely any reduction
2. ρ = 0.5: √0.75 = 0.866
3. ρ = 0: √0.5 = 0.707, about 29% less
4. ρ = −0.3: √0.35 = 0.592

Same profit, smaller swings. The ratio of profit to risk improves, and it is about the only thing in trading you get for free. You don’t need a more accurate signal at all.

Simulation: two systems with a small edge

We generated synthetic daily results in node, not a real market.

1. Each system makes +0.03% of starting capital per day on average, with a daily swing of 1%. The edge is tiny: losing days are almost as common as winning days
2. Run for 1,000 trading days (about 4 years), measured as % of starting capital, no compounding
3. The combined portfolio uses half the size of each system, so capital equals one system
4. 4,000 runs for each value of ρ, reusing the same random numbers for every ρ so they compare directly

Medians of 4,000 runs:

	Total profit	Max DD	Profit ÷ DD	Ends positive
Single system	+29.1%	26.3%	1.10	82.5%
Combined, ρ = −0.3	+29.8%	13.1%	2.24	94.9%
Combined, ρ = 0	+29.7%	16.7%	1.76	91.4%
Combined, ρ = 0.5	+29.8%	21.7%	1.34	86.6%
Combined, ρ = 0.9	+29.2%	25.7%	1.14	83.2%

Three things the table shows:

• Profit does not go up. What improves is profit per unit of drawdown, which is close to the idea behind the Calmar ratio. See Sharpe, Sortino and Calmar
• ρ = 0 already helps a lot. Profit ÷ DD rises from 1.10 to 1.76, about 60%, without finding a better system
• ρ = 0.9 is almost a single system. DD falls from 26.3% only to 25.7%. Two highly correlated systems are one system with double the maintenance

Drawdown falls a little faster than the swing (at ρ = 0 the swing is 0.707 times, but DD is 16.7 ÷ 26.3 = 0.635 times), because the average profit stays the same while the noise shrinks, so the drift beats the randomness sooner.

Effective number of bets: count this instead of systems

With N systems at equal risk, where every pair correlates on average at ρ, the number of truly independent bets is:

N_effective = N ÷ (1 + (N − 1) × ρ)

Average ρ	5 systems	10 systems	20 systems	Ceiling (unlimited systems)
0	5.00	10.00	20.00	unlimited
0.3	2.27	2.70	2.99	3.33
0.5	1.67	1.82	1.90	2.00
0.9	1.09	1.10	1.10	1.11

The last column is what most traders don’t know. As N grows, the formula approaches 1 ÷ ρ. If every system correlates at 0.3, even a hundred systems give you no more than 3.3 independent bets. Past a point, adding systems similar to what you have only adds work. The return-to-swing ratio improves by roughly √N_effective; for a multi-market example see Trend following research.

The formula assumes equal risk per system and the same ρ for every pair, so treat it as a rough figure. With unequal risk the real number is lower.

Three axes of diversification: idea, timeframe, market

Two systems can only have low correlation if they make money from different kinds of events.

• Idea. Trend following makes money when price travels far. Mean reversion makes money when price swings back towards an average. A year of straight-line moves and a year of range-bound chop give opposite results. To measure which behaviour a market is showing, see Hurst exponent and Market regime
• Timeframe. The same idea on very different time scales catches waves of different sizes and only partly overlaps
• Market. Trend following on gold and on a stock index depend on different drivers. But five pairs that all share USD are still one dollar bet

We measured this on one synthetic price series of 2,000 bars (up-trends, down-trends and trendless stretches alternating at random). The trend systems use a fast SMA against a slow SMA (above the slow line = hold a Buy, below = hold a Sell). The MR system trades against price when it is more than 1 standard deviation away from SMA 10. Correlations are of per-bar results over 1,799 bars.

Fake diversification: five versions of one idea

A favourite shortcut among EA builders is to take one working system, tweak its parameters into five sets, run them together and call it “a five-system portfolio”. The numbers in the figure say otherwise.

1. Average correlation of the 10 pairs among the five versions = 0.83
2. N_effective = 5 ÷ (1 + 4 × 0.83) = 5 ÷ 4.32 = 1.16 systems
3. Check with swings: the 20/50 system alone swings 1.01% per bar, the average of all five swings 0.94%, only 7% less. If the five were truly independent it would be 1.01 × √(1/5) = 0.45%

Averaging nearby parameters does have one real benefit: you are not betting that the single value you picked is the best one, which lowers the risk of choosing a peak (Overfitting: peak vs plateau). It does not lower market risk. In the year the idea loses, all five versions lose together.

The figure shows something else. MR correlates negatively with the trend group, yet on this synthetic series MR has no edge (it ends at −2.9%), because we never built any mean-reverting behaviour into the price. Adding it to the portfolio does reduce swings, but it also drags profit down.

Low correlation is necessary but not enough. Every system in the portfolio needs an edge of its own first, otherwise you are just paying spread to cancel out the good one.

Why correlation jumps when markets panic

The ρ you measure in normal times is usually lower than on the days that matter most. We simulated two systems over 200,000 days:

• 95% of days are normal: the two systems are independent (ρ = 0) with a swing of 1 unit
• 5% are panic days: swings are 3 times larger and ρ = 0.9, because everything moves on the same driver

What we measured:

1. Measured on normal days only, ρ ≈ 0.00, which looks like perfect diversification
2. Measured over the whole period, ρ = 0.29, matching the formula (0.05 × 0.9 × 9) ÷ (0.95 + 0.05 × 9) = 0.405 ÷ 1.4 = 0.289
3. On system A’s worst 5% of days, system B was also in its own worst 5% 24.7% of the time. If they were truly independent it would be 5%

So even the full-period average ρ hides risk. You also need to check what the other system does on the days one system takes a heavy loss. In real markets the mechanism is margin calls and forced selling at the same moment, explained in Risk-on, risk-off. Several systems holding overnight positions can also all be hit by one weekend gap at once.

Risk allocation: equal lots are not equal risk

A hypothetical example. Trend system T at 0.10 lot swings 150 dollars a day. System M at 0.10 lot swings 50 dollars a day. Both have the same edge quality (average daily profit = 0.03 × swing) and ρ = 0.

Equal lots, 0.10 and 0.10

1. Average profit = 0.03 × 150 + 0.03 × 50 = 4.5 + 1.5 = 6.00 dollars a day
2. Combined swing = √(150² + 50²) = 158.1 dollars
3. Profit per unit of swing = 6.00 ÷ 158.1 = 0.0379
4. System T drives 150² ÷ (150² + 50²) = 90% of the portfolio’s variance. This portfolio is system T with decorations

Similar risk, T at 0.07 lot, M at 0.22 lot

1. Swing of T = 105, M = 110 dollars
2. Average profit = 0.03 × (105 + 110) = 6.45 dollars a day
3. Combined swing = √(105² + 110²) = 152.1 dollars
4. Profit per unit of swing = 0.0424, 12% better than equal lots, and the portfolio swings less too

To find each system’s swing from test results, use the standard deviation of daily profit and loss from the report. A simpler way is to set lots so each system’s maximum equity drawdown is about the same (Reading drawdown). If you want more risk, scale the whole set by the same factor. Don’t push one system alone.

Rebalancing: pulling the weights back

If each system risks a % of the same account’s total equity, the weights adjust themselves continuously and you don’t need to do anything. The problem appears when you give each system its own pot of money.

Example: split 10,000 dollars, 5,000 per system. After a quarter, A is at 7,000 (+40%) and B at 4,500 (−10%).

1. Total 11,500. A’s share = 7,000 ÷ 11,500 = 60.9%, B = 39.1%
2. Left alone, the portfolio slowly becomes system A on its own and the benefit of diversification fades
3. Rebalance back to 5,750 and 5,750

Common rules are rebalancing on a calendar (for example every quarter) or when the weights drift past a threshold set in advance (for example more than 10 percentage points). These are rough guidelines, not optimal values.

A caution: rebalancing moves money from the winner to the loser. That is fine while the loss is within what testing predicted, not when the system is broken. Set a retirement rule in advance, for example drawdown beyond the range Monte Carlo gave you. Don’t decide from how a losing streak feels, and don’t confuse this with switching a system on and off by its equity, which is a separate topic (Equity curve trading).

What we saw in our own backtests

In our backtests on gold (XAUUSD), January 2022 to July 2026, run separately month by month for 55 months, each month starting from a 3,000-dollar deposit. 2026 used real tick data; 2022–2025 used ticks generated from 1-minute bars. We compared two systems with very different logic: one holds trades for several days, the other uses a near target and closes quickly.

• Correlation of monthly results = 0.074, effectively zero
• Positive months: 60.0% for the long-hold system, 81.8% for the quick-close system
• In the 22 months the long-hold system lost money (−3,121 dollars in total), the quick-close system was positive in 20 of them and made +1,297, offsetting about 42%
• Run together, positive months rose from 60.0% to 67.3%, and the worst month improved from −487 to −406 dollars

Limits you should know: cutting results into months penalises the system that holds for days, because open trades are cut off at month end. The two systems’ lots were not set to equal risk. Both trade one market, so this is not diversification across markets. And the older years’ ticks are not real ticks (tick data quality). Our five-system results are in the correlation article. These are backtest results. Past results do not predict future results, and live results are usually lower (why live differs from backtest).

Putting it to work in 6 steps

1. Make each system pass on its own first, with enough trades (how many trades you need). A system that fails does not become good by joining a portfolio
2. Export daily results for every system over the same period, with 0 on days without trades
3. Compute correlation for every pair with CORREL in a spreadsheet. If a system is above about 0.7 with one you already run, keep only one of them
4. Look at the bad days. Filter each system’s worst 5–10% of days and see what the other system did
5. Set lots for similar risk, then run the combined set in the tester. Compare the combined drawdown with the sum of the individual drawdowns. If they are close, you are barely diversified
6. Cap total open risk for the whole account, assuming every system can lose at once on a panic day. Then fix your rebalancing schedule and retirement rules in advance

Common mistakes

• Counting systems instead of independent bets. Five systems at ρ 0.8 are about one system. Check N ÷ (1 + (N − 1)ρ) every time
• Diversifying by parameters. Nudging a line setting and calling it a new system only reduces the risk of picking the wrong value, not market risk
• Adding a losing system because its ρ is negative. It does cut swings, but it cuts profit too
• Same lot for every system. The system with the biggest swings runs the whole portfolio
• Trusting ρ from a calm period. Re-measure regularly, and plan the worst case as if ρ were close to 1
• Rebalancing into a system that is already broken. Set retirement rules before you start, not while you are losing