Skip to content

The Stock-Bond Correlation (Czasonis, Kritzman, Turkington): summary and key takeaways

The authors argue that extrapolating the correlation of monthly stock and bond returns is an unreliable way to forecast the correlation of five-year returns. They build a forecast from a single-period correlation measure, four fundamental predictors, factor paths and relevance-filtered regression. Each step raises the fit with later actual correlations, and the full model scores best.

Czasonis, M.; Kritzman, M.; Turkington, D. · 2021 · The Journal of Portfolio ManagementSummary updated 20 November 2026
Published2021Journal article
Asset focusStocks+BondsStock-bond correlation
TypeResearchStock-Bond Correlation
Reading time4 minthis summary

The Stock-Bond Correlation (Czasonis, Kritzman, Turkington) at a glance

The question
How can investors forecast the correlation of longer-horizon stock and bond returns more reliably than by extrapolating the correlation of recent monthly returns?
Dataset & period
US stocks (S&P 500; Ibbotson Stock Index before 1973) and US long Treasuries (Bloomberg US Long Treasury Index; Ibbotson Treasury Bond Index before 1973). Correlation history January 1926 to November 2019, monthly returns. Predictions start January 1930, target five-year single-period correlations through December 2018, updated annually. Number of possible offsets in the simulation test: 82.
Method
The paper defines a single-period correlation for each five-year window. It is the product of the stock and bond z-scores for that window, divided by the average of their squared z-scores. The z-scores use full-sample means and standard deviations of five-year returns. Five models forecast the next window: naive extrapolation of the trailing monthly correlation, a regression on that trailing correlation, a regression on four fundamentals (industrial production growth, inflation, relative stock/bond yield, relative yield volatility), the same regression filtered to the most relevant past observations, and a version that uses five-period paths of growth and inflation with the same filter. Relevance combines similarity to current conditions with how far past observations sit from their averages. Forecasts use the 25% most relevant observations.

Key findings of The Stock-Bond Correlation (Czasonis, Kritzman, Turkington)

  • The stock-bond correlation depends on the return interval: 0.07 for monthly, 0.11 for annual, 0.17 for 3-year and 0.21 for 5-year returns, January 1926 to November 2019.
  • Naive extrapolation of the trailing monthly correlation and a regression on it tie as the least reliable models. Both show a correlation of 0.23 with the actual subsequent single-period correlation.
  • Adding the four fundamental variables lifts that correlation to 0.34. Filtering for relevance lifts it to 0.57. Adding paths of industrial production and inflation gives 0.63.
  • Henriksson-Merton scores for predicting the sign of the correlation level run from 1.06 (extrapolation) to 1.29 (full model). For the sign of the change, the scores run from 1.13 to 1.41.
  • The authors report p-values of 0.29 and 0.00 at the two ends of the level-sign comparison, and 0.19 and 0.01 for the change-sign comparison.
  • The simulation test gives a p-value of 0.06 for the filtered fundamentals model, against 0.27 for trailing correlation and 0.44 for unfiltered fundamentals.

Robustness, as reported in the paper

  • Because adjusted R-squared cannot be used with partial sample regression, the authors shift the dependent series against the predictors 82 times to build a distribution of chance correlations for each model.
  • In that test, only five random correlations exceeded the actual result for the filtered fundamentals model (p-value 0.06).
  • The paper notes that the paths-and-filtered model and the unfiltered fundamentals model beat trailing correlation but are less statistically significant (p-values 0.37 and 0.44).
  • The Henriksson-Merton sign tests are offered as a check that is less affected by small differences around zero correlation.

Caveats and limits

  • The study covers one country pair (US stocks and long Treasuries) and one horizon (five years). The authors say transfer to other asset pairs is an empirical question.
  • The evaluation rests on a single US history (1926 to 2019) and 82 possible offsets in the simulation test, and several p-values are well above conventional thresholds (inferred).
  • The paper does not report transaction costs or an investable strategy built on the forecasts (inferred).
  • The 25% relevance cutoff, the choice of four fundamentals and the path length are design choices. The paper does not report sensitivity to them (inferred).

Items marked (inferred) are our reading of the design, not statements by the authors.

Why it matters

Portfolio construction, hedging and risk models often assume a stock-bond correlation estimated from short-interval data. The paper shows that this input can differ a lot from the correlation of longer-horizon returns. It also gives a framework for conditioning the forecast on economic and valuation state.

Questions about this paper

Why not use the monthly stock-bond correlation to forecast longer-horizon correlation?

The paper shows the measured correlation changes with the return interval, from 0.07 for monthly to 0.21 for 5-year returns. It calls extrapolating the monthly figure decidedly unreliable.

What is a single-period correlation?

It measures how stocks and bonds co-moved over one specific five-year window, relative to their full-sample average returns and volatility. A weighted average of these values across the sample gives the usual Pearson correlation.

Which predictors does the paper use?

Industrial production growth, inflation, the relative yield of stocks and bonds, and the relative volatility of stock and bond yields. Growth and inflation enter as five-period paths in the final model.

Where can I read the original paper?

The original is at its publisher: https://doi.org/10.3905/jpm.2020.1.195. We link to the paper and do not host the PDF.

Follow new quant research

Download MacroFinance market data as CSV or Excel, or pull it by API, and test the ideas in these papers yourself. New summaries are added as we review them.

Create a free account
Summaries are written by MacroFinance from the published papers, in our own words, for information only. They are not investment advice and may omit detail. Papers are not hosted here: each page links to the original at its publisher.