The Correlation stage is diversification’s gatekeeper. After the filters have chosen the good strategies, this stage compares them pairwise and excludes those too similar to each other, because a portfolio made of strategies moving in unison is, in effect, one single bet disguised as many.
In simple terms: two strategies can both be excellent and, taken together, add very little to each other — because they win and lose at the same moments. When the bad hit comes, they take it together, and the portfolio’s drawdown gets deeper instead of flatter. The Correlation stage identifies these redundant pairs and keeps only one, the best according to a criterion you set. The result is a portfolio whose components are more independent, and therefore an overall equity smoother than that of the individual strategies.
Two traits make it distinctive. First: it’s a single final stage, not a filter in the list — because decorrelation, unlike a quality gate, is path-dependent (each exclusion changes the subsequent ones). Second: it excludes only positive correlations above the threshold. Two strategies moving in opposite directions are pure diversification, not risk: the stage never touches them.
An important premise. QANTHOS is an analysis, validation and discovery tool, not a financial advisor. The method, threshold and parameters described in this chapter are configurable levers: they establish how the stage measures and excludes, not how much correlation is “acceptable” for you. That judgement depends on your pool, your instruments and your diversification goals: you are always the final judge.
1. Why decorrelation is a stage of its own
It’s worth understanding why the Correlation stage doesn’t live in the ordered filter list alongside Profit Factor, Sortino and the rest. The reason is a deep difference in nature.
A classic scalar filter — “keep only those with Profit Factor above 1.0” — judges each strategy on its own. Its verdict on a single strategy doesn’t depend on which other strategies survive: whether you place it before or after another quality gate, the outcome for that strategy is the same. These are independent decisions, and their order doesn’t change the final result.
Decorrelation is different. It’s greedy and path-dependent: the decision on a strategy depends on which others have already been kept. Excluding a strategy because it’s too correlated with another changes the set of survivors, and therefore changes all subsequent evaluations — because the next pair to examine, and the outcome of that comparison, depend on who’s still in play at that point. A mechanism like this doesn’t behave like an independent filter: it wouldn’t make sense to “interleave” it among the other filters, because its result would change depending on where you place it.
That’s why QANTHOS treats it as a single final pass, applied once after all ordered filters have done their work. Filters choose who’s good; the Correlation stage, among the good ones remaining, chooses who’s also different from the others. Keeping the two kinds of decision separate — “per strategy” on one side, “on the set” on the other — is what makes the system’s behaviour predictable.
2. The single rule: positive correlations only
The stage’s core is a single rule, and it has an important trait: it acts in one direction only. The stage excludes a pair of strategies only when their similarity is positive and exceeds the threshold you’ve set. Never on anti-correlated pairs.
The reason is simple and deep. Two positively correlated strategies reinforce each other in gains but also in losses: together they amplify movements, and when one suffers so does the other. They’re redundant, and they concentrate risk. Two anti-correlated strategies, instead, do the opposite: when one loses, the other tends to gain, and their ups and downs offset each other. That’s exactly what you look for in a portfolio — diversification, not risk. Excluding them would be counterproductive.
That’s why the stage doesn’t look at the correlation’s absolute value, but at its sign: it filters pairs too similar, leaves alone (indeed, values) those that balance each other. It’s a design choice consistent with the whole module’s purpose — reducing redundancy, not variety.
In practice. If you set the threshold to
0.70, two strategies with correlation+0.82are candidates for exclusion (one of the two drops out); two strategies with correlation−0.60, however “far” from zero, are never touched. The sign matters more than the distance.
3. The three measurement methods
The stage offers three methods, mutually exclusive: only one is chosen per generation. They measure a pair’s “similarity” in three different ways, each sensitive to a distinct aspect of shared risk.
| Method | What it measures | When it’s particularly useful |
|---|---|---|
| Classic Correlation | The pair’s average equity correlation: how much their returns move together over time. The most direct, intuitive measure of “similarity”. | As the general first-choice decorrelation. Captures overall behavioural overlap. |
| Co-Loss | The frequency with which the two strategies lose in the same periods. Doesn’t look at the shape of returns, but at the coincidence of losses. | When you care about drawdown more than average correlation: two strategies can have modest correlation yet tend to lose together. Co-Loss captures exactly this. |
| Stress Correlation | The correlation computed during stress/drawdown phases, ignoring calm periods. | When you fear diversification “vanishes” in the worst moments — the classic phenomenon where markets and strategies correlate exactly when you’d want them to stay independent. |
Why three methods and not one. Average correlation (Classic) is a good summary, but it’s an average: it can hide the fact that two strategies, though different in normal times, crash together in critical moments. Co-Loss and Stress Correlation are two stricter lenses, designed for tail risk — what really matters when the portfolio is under pressure. Which lens to use depends on what you want to protect; there’s no absolutely “right” choice.
A method may appear disabled in the dialog if the data needed to compute it isn’t available for the current pool: in that case QANTHOS flags it and leaves selectable only the computable methods.
4. Where it acts in the chain
The Correlation stage sits after all ordered filters and before the Portfolio Rebalancer (which in turn precedes Money Management). The order has a logic: first filters choose the quality strategies, then Correlation removes the redundant ones, then the Rebalancer looks at concentration on tickers and asset classes, finally Money Management sizes the survivors.
Like the other stages, it acts at every rotation: from rotation to rotation the pool changes, and the stage recalculates correlations and exclusions from scratch on the current composition. Its effect, therefore, isn’t about a single moment but shapes the entire portfolio equity over time.
5. The configuration dialog
The stage opens from the 🔗 Correlation button in the Portfolio Rules header. Unlike other modules it has no separate enable switch: it’s activated by choosing a method and deactivated by selecting the — Disabled — option in the menu. The dialog is organized into three parts.
5.1 Method
The Decorrelation method menu is where you choose the single method to apply: — Disabled —, Classic Correlation, Co-Loss (simultaneous losses) or Stress Correlation. It’s a single choice: the three methods don’t combine.
5.2 Settings (common to all methods)
| Control | Meaning |
|---|---|
| Maximum threshold | The threshold beyond which a pair is “too similar”. Its meaning depends on the method: for Classic it’s the maximum allowed average correlation; for Co-Loss it’s the maximum joint-loss frequency (e.g. 0.20 = 20%); for Stress it’s the maximum stress correlation. In all cases it applies only to positive values. The field’s label changes automatically to remind you which quantity you’re limiting. |
| Ranking metric (who to keep) | The metric the stage uses to decide, in an over-correlated pair, which strategy to keep: the best by this metric survives, the other drops. Options: Net Profit (default), Sharpe Ratio, Profit Factor, Sortino Ratio, Calmar Ratio, Avg Trade. |
| Lookback | The window in months over which to compute correlation, up to the rotation’s evaluation date (default 24 months). 0 means “all available history up to that date” — never beyond, so there’s no look-ahead. |
| Sampling frequency | Which returns to sample the calculation on: daily (granular, noisier), weekly (more balanced, default), monthly (smoother but needs more history). |
On lookback and look-ahead. The stage always looks only at the past relative to each rotation’s evaluation date: even with lookback set to 0 (“all history”), it never peeks beyond the current date. This is a guarantee of walk-forward correctness — a rotation’s decorrelation is based exclusively on what would have been known at that moment.
5.3 Classic Correlation — advanced
This box appears only when the selected method is Classic. It gathers two fine-grained parameters governing how and when classic correlation is computed. For most uses the defaults are fine; they’re here for those who want detailed control.
| Control | Meaning |
|---|---|
| Calculation method | How to compute correlation: sign (sign correlation: measures agreement on the direction of movements, robust to outliers; QANTHOS’s default), pearson (classic linear correlation), spearman (rank correlation, captures non-linear monotonic relationships). |
| Mode (when) | When to measure: active_days (only periods when both strategies were active — default), any_active (periods when at least one was active), calendar (all calendar periods, including “idle” ones), downside (concentrates the measure on negative periods). |
Why “sign” and “active_days” as defaults. Sign correlation looks at whether two strategies move in the same direction, more than by how much: it’s less sensitive to a single anomalous trade that would inflate a linear correlation. active_days mode compares the two strategies only when both are actually in the market, preventing long shared idle periods from producing an “apparent” correlation that doesn’t reflect real behaviour. Together they tend to give a more stable, honest measure of similarity. They remain choices, though: the other methods and modes are there for those with different needs.
6. How it chooses who to keep: the greedy process
When a pair exceeds the threshold, one of the two strategies must drop out. Which one? The Ranking metric decides: the one with the better value by that metric survives, the other drops. If the ranking metric is Net Profit, in an over-correlated pair the more profitable one stays and the less profitable one is excluded.
The process is greedy and iterative: the stage progressively builds the set of survivors, and for each strategy checks whether it’s too correlated with one already kept. If so, it excludes it — also recording the “partner”, i.e. the already-selected strategy that caused its exclusion. Each exclusion changes the set, and with it subsequent evaluations: exactly the path-dependent nature mentioned at the start. The net result is that, among a group of strategies all similar to each other, the best one by the chosen criterion survives, and the others are let go.
The mechanism is deterministic: given the same input and configuration, it always produces the same set of survivors.
7. Correlation stage vs Money Management’s Correlation Penalty: two different things
There’s a point that’s easy to confuse, because the word “correlation” appears in two places in the Builder with opposite operational meanings. Worth fixing.
The Correlation stage (this chapter) excludes: it decides who’s in the portfolio, removing strategies too similar. It’s a binary decision — in or out.
Money Management‘s adjustments — in particular the Correlation Penalty tab, with its Classic and Co-Loss systems — exclude no one: they resize. They work after selection, on already-chosen strategies, and modulate their size (more contracts to decorrelated ones, fewer to correlated ones). It’s a dosing decision — how much weight, not in/out.
| Correlation stage | Correlation Penalty (Money Management) | |
|---|---|---|
| Effect | Excludes strategies | Resizes |
| Question | Who’s in the portfolio? | With how much weight, among those in? |
| Position in the chain | After filters, before the Rebalancer | Last, on sizing |
The two tools aren’t alternatives: they’re often used together. First the stage removes the most marked redundancies (who’s in), then Money Management doses the survivors’ sizes accounting for residual correlations (with how much weight). Interestingly both offer a “Classic” and a “Co-Loss” variant: the same measurement idea, but one excludes and the other resizes. Money Management is covered in its own dedicated sub-chapter.
8. Common issues and notes
I selected a method but it excludes nothing. Several possible causes, all legitimate: no pair exceeds the threshold (try lowering it); pairs above threshold are anti-correlated (and therefore, correctly, aren’t touched — only positive counts); or the pool is already well decorrelated. At the end of generation QANTHOS distinguishes “wasn’t needed” from “wasn’t computable”, so you know which case you’re in.
The method I want is disabled in the menu. The data needed for that method isn’t available for the current pool. Choose a computable method, or check whether strategies have enough history within the lookback window.
I change the Ranking metric but the number of excluded strategies doesn’t change. That’s normal: the ranking metric doesn’t decide how many pairs exceed the threshold (that depends on the method and threshold), only which of the two drops in each pair. It changes who survives, not how many survive.
The stage feels too aggressive / too lax. The lever is the threshold. Lower = stricter (more pairs considered “too similar”); higher = more permissive. Remember the threshold’s numeric meaning depends on the method (average correlation for Classic, co-loss frequency for Co-Loss, stress correlation for Stress).
9. An example workflow
An exploration path, not a recipe. The right threshold and method depend on your pool and what you want to protect.
- Open 🔗 Correlation from the Portfolio Rules header.
- Choose a method in Method (Classic for general decorrelation; Co-Loss or Stress if tail risk worries you).
- Set the threshold in Settings, remembering its meaning depends on the method.
- Choose the Ranking metric: in a pair too similar, the best by this metric survives.
- Adjust Lookback and Sampling frequency based on rotation frequency and how much history your strategies have.
- (Classic only) If you want, fine-tune Calculation method and Mode in the advanced box.
- GENERATE PORTFOLIO and observe, in the Equity tab and the analysis tools, decorrelation’s effect on equity smoothness and drawdown.
The Correlation stage translates a common-sense intuition into an automatic rule: many strategies doing the same thing aren’t diversification. But you choose the threshold and the method, and the real effect on portfolio diversification should always be verified — with Walk-Forward Analysis, stress testing and drawdown observation — before considering it settled.
10. To learn more
- Portfolio Rules — the rule chain and where the Correlation stage fits.
- Money Management — the Correlation Penalty and Co-Loss adjustments that resize (distinct from this stage, which excludes).
- Portfolio Rebalancer — concentration balancing, applied right after the Correlation stage.
- Custom Index — for building a custom composite ranking metric.
- Tools Tab — reading the stages’ rotation-by-rotation impact (Equity, Filters Analysis).