Research cut-off: August 14, 2026. Price history ends on August 13, 2026. Correlations, volatility, and market regimes can change after publication.
Gold, U.S. equities, and Bitcoin are usually introduced as if each has one job. Stocks compound growth. Gold protects capital. Bitcoin sits outside the old system and diversifies both. The neat version of that story breaks down as soon as the market changes character.
Weekly data since 2017 gives a less comfortable answer. All three pairwise correlations rose when U.S. equities entered the bear regime used in this report. Gold still carried far less equity beta and much lower volatility than Bitcoin, but its correlation with stocks did not turn negative. Bitcoin remained the more aggressive risk asset and fell alongside stocks more often when the equity tape was weak.
Diversification is conditional. A portfolio built from one full-sample correlation matrix can look sturdier on paper than it behaves under stress.
Executive summary
- The correlation between weekly SPY and GLD returns rose from 0.15 in bull regimes to 0.36 in bear regimes. Gold was still the lower-beta asset, but it was not a consistent negative-correlation hedge.
- The SPY/BTC correlation increased from 0.20 to 0.28. During bear-regime weeks in which SPY fell, Bitcoin also fell 68 percent of the time, compared with 55 percent in bull regimes.
- The largest regime change was between GLD and Bitcoin. Their correlation rose from 0.07 to 0.38, evidence that a common liquidity or dollar shock can pull two assets with very different stories in the same direction.
- Volatility changed the portfolio meaning of those correlations. Annualized weekly volatility in bear regimes was about 29 percent for SPY, 18 percent for GLD, and 74 percent for Bitcoin. Equal portfolio weights would not create equal risk contributions.
The test uses returns, not price lines
Two assets can both rise over ten years and produce a high correlation between their price levels even if their short-term returns have little relationship. That is a classic route to spurious correlation. This report therefore measures correlation between weekly log returns, not between the three price lines.
SPY is the investable proxy for U.S. large-cap equities. GLD is the gold proxy. The SPDR Gold Trust is designed to reflect the price of gold bullion, less expenses, and uses the LBMA Gold Price PM as its benchmark. Bitcoin uses Coin Metrics' daily USD price metric, which is built from its rules-based reference-rate methodology.
The price series are Nasdaq historical closes for SPY, Nasdaq historical closes for GLD, and Coin Metrics PriceUSD for Bitcoin. Each weekly observation uses the last date on which all three series have a value. Returns are then measured from one common weekly close to the next.
The sample contains 482 weekly return observations from May 26, 2017 through August 13, 2026. A week is assigned to the bull regime when the previous week's SPY close is at or above its trailing 40-week simple moving average. It is assigned to the bear regime when the previous close is below that average. Using the prior close avoids using future information to label the current return.
This is a trend-regime definition, not the newspaper definition of a bear market as a 20 percent drawdown. It produced 384 bull-regime weeks and 98 bear-regime weeks. The label describes the U.S. equity tape, not a universal state shared by every asset.
Correlation rises when the equity tape weakens
The regime split produces a clean result. Every pair became more positively correlated in bear regimes.
The SPY/GLD correlation increased by 0.21, from 0.15 to 0.36. SPY/BTC rose by a smaller 0.08, from 0.20 to 0.28. GLD/BTC increased by 0.31, from almost no relationship at 0.07 to a moderate positive relationship at 0.38.
These are descriptive statistics, not proof that a weak stock market causes gold and Bitcoin to move together. The regime can coincide with tighter liquidity, a stronger or weaker dollar, changing real yields, forced deleveraging, geopolitical stress, or several shocks at once. Correlation records the shared movement. It does not identify the driver.
For a portfolio, the timing is the point. A negative stock week is when investors want diversification to do more work. In the sample, neither GLD nor Bitcoin offered a stable negative relationship with SPY during bear regimes. Diversification weakened, even though the three assets did not become interchangeable.
Rolling lines show that no correlation is permanent
A bull/bear average compresses a moving relationship into two numbers. The 26-week rolling series shows how much information that average hides.
Across the full chart, SPY/GLD moved between about -0.66 and +0.68. SPY/BTC ranged from roughly -0.56 to +0.66. GLD/BTC ranged from about -0.31 to +0.68. None of the three lines stayed near a single long-run value.
The latest 26-week window ended with correlations of approximately 0.68 for SPY/GLD, 0.56 for SPY/BTC, and 0.31 for GLD/BTC. Those readings are snapshots, not forecasts. Treated as permanent, a six-month reading of 0.68 is alarming. Treated as a forecast, it is fragile. The window reacts quickly to a common shock and can reverse just as quickly when that shock fades.
This helps reconcile older and newer research. A widely cited 2018 NBER study found cryptocurrency returns distinct from stocks, currencies, and precious metals. That result described an earlier market structure. Later work, including a 2024 study of Bitcoin, gold, and U.S. equity volatility around COVID-19, found that conditional correlations changed during stress. Both can be true. Correlation is a property of a sample and a market regime, not an asset's permanent identity.
Bitcoin's connection to traditional portfolio flows has also grown. The SEC approved the listing and trading of spot Bitcoin ETP shares in January 2024. That decision gave brokerage and advisory accounts a regulated access route. It does not prove that ETFs caused the correlations in this report, but it made it easier for Bitcoin exposure to enter the same allocation, rebalancing, and risk-control process as equities and gold.
Bitcoin absorbs more downside; gold keeps the lower beta
The same correlation can have very different portfolio consequences when the assets have different volatility and beta. Bitcoin and gold illustrate the distinction.
Bitcoin's correlation rises, but its beta falls
The SPY/BTC scatter is positive in both regimes. The bear-regime relationship is somewhat tighter, with correlation rising from 0.20 to 0.28. Yet Bitcoin's estimated equity beta falls from 0.99 to 0.70.
That is not evidence that Bitcoin becomes safer in bear markets. Beta depends on the relative volatility of the two assets. SPY volatility more than doubled in the bear regime, while Bitcoin volatility rose much less. A modestly tighter relationship divided by a much larger equity variance produces a lower slope.
The direction of down weeks is more intuitive. In bull regimes, Bitcoin fell in 55 percent of weeks when SPY fell. In bear regimes, that share increased to 68 percent. Bitcoin did not move with stocks every week, but the odds of concurrent losses became materially worse when the equity trend was already weak.
Gold's protection came from low beta, not negative weekly correlation
Gold's equity beta was 0.16 in bull regimes and 0.21 in bear regimes. Those are much lower than Bitcoin's estimates. Gold also fell in 48 percent of negative-SPY weeks during bear regimes, close to a coin toss and well below Bitcoin's 68 percent.
Still, the SPY/GLD correlation was +0.36 in bear regimes, not negative. Gold can protect a portfolio through lower beta, lower volatility, a different source of demand, or gains over a full drawdown window. It does not need to rise on every bad stock week to diversify. The data does not support the stronger claim that gold is reliable one-week crash insurance in every bear regime.
Volatility changes what a correlation number means
Correlation is scale-free. Portfolio losses are not. Bitcoin's annualized weekly volatility was about 66 percent in bull regimes and 74 percent in bear regimes. GLD moved from 15 percent to 18 percent. SPY moved from 13 percent to 29 percent.
An equal-dollar portfolio would therefore receive far more risk from Bitcoin than from either SPY or GLD. During the bear regime, Bitcoin's volatility was about 2.5 times SPY's and more than four times GLD's. Even a 0.28 correlation can transmit a large amount of portfolio stress when the asset carrying that correlation is exceptionally volatile.
A correlation matrix needs a volatility estimate beside it. Correlation describes the direction and consistency of co-movement. Volatility describes the size of each move. Covariance and portfolio risk depend on both.
A second regime definition keeps the same result
Any bull/bear test is partly a test of its definition. The 40-week moving average is transparent and uses no future data, but another reasonable rule may classify some weeks differently.
As a robustness check, the sample was relabeled using the prior 26-week SPY return. A non-negative trailing return defines the bull regime; a negative trailing return defines the bear regime. This produces 376 bull weeks and 106 bear weeks.
| Pair | Bull correlation | Bear correlation | Bear minus bull |
|---|---|---|---|
| SPY / GLD | 0.15 | 0.34 | +0.19 |
| SPY / BTC | 0.21 | 0.26 | +0.05 |
| GLD / BTC | 0.07 | 0.37 | +0.30 |
The exact values move, but the ordering does not. All three correlations are higher in the alternative bear regime, and the largest change remains the GLD/BTC pair. The baseline finding is therefore not an artifact of one moving-average threshold.
Several limitations remain. Nasdaq closes are market prices, not total returns, so SPY dividends are excluded. GLD is an ETF proxy rather than licensed LBMA spot history. Bitcoin trades continuously, while SPY and GLD trade on U.S. exchange hours. A common calendar reduces timing mismatch but cannot remove it. Pearson correlation captures linear co-movement and can miss tail dependence, lead-lag effects, and nonlinear relationships. The sample begins in 2017 after the 40-week lookback, so it says little about Bitcoin's earlier market structure.
Most important, this is not a causal model and not an investment recommendation. It does not estimate expected returns, inflation sensitivity, real-rate sensitivity, transaction costs, taxes, or how a particular allocation would rebalance through a drawdown.
Portfolio implications and open questions
A regime-aware process does not need a complicated forecast. It needs fewer permanent labels.
- Estimate correlation on the horizon you actually hold. Daily, weekly, and monthly data answer different questions. A six-month rolling window is useful for monitoring but too unstable to anchor a strategic allocation on its own.
- Stress correlation and volatility together. Replacing a bull-regime correlation matrix with a bear-regime matrix is only half the test. The volatility inputs need to change at the same time.
- Treat Bitcoin as an independent source of return rather than guaranteed insurance. Its long-run monetary case may differ from equities, yet its short-run losses can still arrive with stock losses.
- Use gold as a lower-beta diversifier without expecting it to rise in every negative stock week. Its portfolio value can come from smaller moves and different long-window behavior even when contemporaneous correlation is positive.
- Size by risk contribution. A portfolio with one-third in each asset is not balanced when one sleeve carries four times another's volatility.
Open questions
The ETF era is still short. A longer post-2024 sample is needed to know whether Bitcoin's higher equity linkage is structural or another rolling phase. Stress also has different causes. An inflation shock, a real-rate shock, a funding squeeze, and a geopolitical event should not be expected to produce the same gold or Bitcoin response. Future work should separate those shock types and test tail dependence, not only average correlation.
The data supports a narrower conclusion. Gold, U.S. equities, and Bitcoin did not keep fixed relationships across the sample. Correlation rose when the U.S. equity trend weakened, while volatility changed by very different amounts. Diversification still existed, but it was least reliable when a static model looked most reassuring.
Sources
This report uses public market data and published research available at the research cut-off. All correlation, beta, volatility, and regime calculations are by InvisibleHill.
- Nasdaq, SPY historical data
- Nasdaq, GLD historical data
- SPDR Gold Shares, fund objective and benchmark
- Coin Metrics, PriceUSD definition
- Coin Metrics, prices methodology
- Liu and Tsyvinski, Risks and Returns of Cryptocurrency, NBER Working Paper 24877
- Elsayed and coauthors, The nexus between the volatility of Bitcoin, gold, and American stock markets during the COVID-19 pandemic
- U.S. Securities and Exchange Commission, statement on spot Bitcoin ETP approval