When did stocks and bonds stop moving apart?

The 60/40 portfolio rests on one statistic: the correlation between equity and long bond returns. For two decades it was negative, and every allocation built on that fact worked. Measured year by year, the correlation has a date on which it stopped.

Measured, 2026-09-04

Correlation of daily returns between SPY.US and TLT.US, the 20-year Treasury fund, on adjusted closes:

Read down the column. Through 2020 the two moved apart, and moved apart most in the crisis month, which is what a hedge is for. In 2021 the relationship weakened; in 2022 it crossed zero, and in 2022 both fell — the S&P 500 fund by 18.2%, the Treasury fund by more than 30% — and a 60/40 portfolio had its worst year since 2008 with nothing in it that went up. Since then the correlation has stayed positive and in 2026 it has been the highest on the table.

Why it flipped

The negative correlation was a feature of a regime, not a law. When the dominant risk is growth — a recession, a crash — bad news for shares is good news for bonds, because it brings rate cuts, and the two move apart. When the dominant risk is inflation, bad news is bad for both: higher inflation means higher rates, which is a loss on the bond and a lower multiple on the share. 2022 was the first inflation year in the sample, and the correlation did what the mechanism predicts. The 1970s, before this sample, were positive-correlation years for the same reason.

What the instability means

A correlation is an average over a window and a regime, and the "−0.31 full period" figure describes no year in the table — it is the mean of a negative decade and a positive half-decade. Any risk model that uses the full-period number is using a number that has not been true since 2021.

The direction of the error matters. A model that assumes −0.3 when the truth is +0.3 understates a 60/40 portfolio's volatility by a wide margin, because the diversification it counts on is not there. The allocation course's reference mixes lesson argued 60/40 on the old figure; this table is why the argument is now conditional.

The rolling-windows lesson has the mirror image: the same relationship measured over 60, 250 and 1,000 days on the same date gives three answers, and none of them is wrong.

In the data

Two pulls — /eod/SPY.US?from=2006-09-01&to=2026-09-03&fmt=json and the same for TLT.US — joined on date, converted to returns, and correlated within each calendar year. A day present in one series and not the other must be dropped, not filled; the computing-returns lesson is about that join.

Try it now

1. A year of daily returns is 250 pairs, too many for a page; twelve monthly ones are not. Here are both funds' thirteen month-end closes for 2022. Compute the twelve monthly returns of each and their correlation, and write it beside the table's daily +0.08. Measured the same way on 28 September 2026, 2019's twelve monthly returns correlated at −0.60 against the table's daily −0.46. Say what the monthly figures agree with the daily ones about, and why they need not agree on the size.

2. The two funds over the window in which the flip happened:

Measure both over calendar 2022. Two negative numbers, on the two assets that were supposed to hedge each other. 3. Your call: which single macro fact would you watch to guess whether the correlation goes back below zero, and why?