Bivar Capital — Quantitative Research

Canada & USA MomentumCorrelation, crisis behavior & combined portfolio

How correlated are our Canada TSX and USA momentum strategies — overall, and specifically during the six largest drawdowns since 2000? Monthly returns, both net of costs and regime, 2000–2026.

I. Headline Number

0.346
Pearson r, full sample
316
Months (2000–01 to 2026–04)
p<0.001
Statistically significant
[0.25, 0.44]
95% confidence interval

Positive and real — not noise — but moderate. Not close to 1, and not close to 0 either: enough shared exposure to broad market and commodity cycles that the two aren't independent bets, but enough daylight that combining them does real diversification work (Section IV).

It moves a lot. Rolling 12-month correlation ranges from −0.73 to +0.85 (mean 0.34, std 0.32) across the sample, and is outright negative in 15.4% of 12-month windows. A single "the correlation is 0.35" number hides a genuinely unstable relationship — which is exactly why the rest of this article looks at specific periods instead of stopping here.

II. A Mechanical Artifact to Rule Out First

Both strategies carry a regime filter that moves them to 100% cash when their local index is below its moving average. During a drawdown, both books are very often sitting in cash at the same time — two flat, near-zero return streams will show high correlation with each other almost by construction, whether or not the underlying markets are actually moving together. Before trusting any crisis-period correlation number, we recompute it on "active" months only — months where at least one strategy had a real (>1%) move, excluding the both-parked-in-cash months that would otherwise inflate the number for free.

SampleNCorrelation
Full sample3160.346
Active months only (excl. both-cash)2530.326

Close enough (0.346 vs 0.326) that the full-sample number isn't meaningfully inflated overall — but as Section III shows, the gap between the two matters a lot more inside individual short crisis windows, where the sample size is small enough for a couple of cash months to swing the number substantially.

III. Six Crises, Two Correlation Numbers Each

Windows are standard, well-known drawdown periods, chosen by calendar date rather than detected algorithmically from either return series (so the choice of window isn't itself curve-fit to either strategy's results).

CrisisMonthsCanada cum.USA cum.Corr (all)Corr (active only)
Dot-com crash (2000–03 to 2002–10)32+32.9%+14.5%0.5490.545
GFC (2007–10 to 2009–03)18−3.5%+13.1%0.2920.319
US downgrade / Euro crisis (2011–05 to 2011–10)6−0.3%−7.5%−0.0990.002
Oil / commodity crash (2015–08 to 2016–02)7−0.6%−11.0%0.6600.783
COVID crash (2020–02 to 2020–04)3−11.1%−7.3%1.000n/a (1 active month)
2022 bear market (2022–01 to 2022–10)10+34.4%−0.5%−0.655−0.868

COVID's "1.000" is not a real number — ignore it. Once cash months are stripped out, only one month in the entire three-month window had both strategies actually invested (Feb 2020). Correlation from a single data point is undefined, not perfect; the raw 1.000 is an artifact of two mostly-flat, mostly-cash series moving together in the one month they were both still invested, not evidence the strategies behave identically in a liquidity panic. We're showing it precisely so we can say this clearly, not to quietly drop the inconvenient case.

2022 is the standout real diversification case. −0.655 raw, and it gets more negative (−0.868), not less, once cash-timing is stripped out — this is genuine anti-correlated performance, not an artifact. Canada's commodity-heavy book was up +34.4% cumulative through a year that hammered US growth/tech names; USA's momentum book, concentrated in whatever was working (semiconductors and related names for parts of this window), was roughly flat. This is exactly the kind of year a combined book is built for.

The oil crash is the opposite lesson: correlation rose when it mattered. 0.660 raw, 0.783 active-only — both strategies were hit by the same 2015–16 global growth scare, just to different degrees (Canada roughly flat at −0.6%, USA down −11.0%). Diversification did not fully show up here. Worth remembering: the 2022 result above is a real pattern, not a guarantee that applies to every risk-off period.

IV. Combined Portfolio: Where 50/50 Actually Sits

Using the full 316-month sample, monthly rebalanced, no re-optimization over time (fixed weights throughout):

AllocationCAGRSharpeMaxDDVol
100% Canada+37.0%1.39−24.5%24.6%
70/30 (Canada/USA)+33.8%1.46−21.6%21.1%
Max-Sharpe (74/26)+34.2%1.46−21.4%21.4%
60/40+32.5%1.43−22.4%20.6%
Min-Variance (53/47)+31.6%1.39−23.0%20.5%
50/50+31.1%1.37−23.3%20.6%
100% USA+23.3%0.79−33.2%25.6%

50/50 is not on the efficient frontier here. 70/30 (Canada-heavy) beats it on every metric simultaneously — higher CAGR, higher Sharpe, and shallower MaxDD. This happens because Canada alone already has a better risk-adjusted profile than USA (Sharpe 1.39 vs 0.79); the math doesn't reward splitting evenly between a stronger and a weaker Sharpe just because the two are diversifying. 50/50 is an intuitive starting point, not an optimized one.

V. Limitations

  1. USA data window is shorter and stale. Sharadar data (the USA backtest's source) is current only through mid-May 2026, so this entire comparison stops at April 2026 — it excludes Canada's -15.6% March 2026 and does not reflect either strategy's most recent months.
  2. April 2026 carries a known USA data anomaly. SNDK, ineligible under the backtest's own 252-day-history rule but not gated in the live signal, drove an unrepresentative +53.7% single-month USA return inside this window — a documented, separate issue (see the USA strategy's own audit notes), not corrected in the series used here.
  3. Different data vendors. Canada sources from EODHD, USA from Sharadar; universe construction, corporate-action handling and data quality standards differ between them in ways that aren't fully reconciled.
  4. Crisis windows are hindsight-selected. Six well-known calendar periods, not an algorithm that scans for the worst joint drawdowns — a systematic worst-case search could turn up a period not covered here.
  5. Fixed historical weights, not a live recommendation. Section IV uses one static allocation across the full 26-year sample. It says where the historical efficient frontier sits, not what to hold today — that's a portfolio-construction decision, not a backtest output.