Idiosyncratic Market Effects in Global Time Series Momentum

summary

The gist

Time series momentum models fit at the individual market level exhibit idiosyncratic forward return characteristics that are distinct from global linear and nonlinear time series momentum factor…

In short

The study compares time series momentum models across individual markets versus global factor models. It found that fitting nonlinear models at the market level captures unique prediction behaviors not seen in global linear or nonlinear models, suggesting this approach can improve out-of-sample performance.

Key concepts

Time Series Momentum Models
These are statistical methods used to forecast future returns based on how much an asset has moved recently. The paper tests how these models perform when applied separately to individual markets rather than as a single global average.
Individual Market Level Fitting
Instead of using one model for all markets, this approach fits separate models for each market. This allows each specific market to show its own unique way of reacting to momentum features, which is different from a single global model.
Kernel Regression
This is a non-parametric statistical technique used in the modeling. It allows the model to create flexible, nonlinear response functions by taking weighted averages of nearby data points, enabling it to capture complex relationships that simple linear models miss.

Terminology used across episodes

This episode discusses

The paper

Idiosyncratic Market Effects in Global Time Series Momentum · Read on arXiv

Andrew Butler, Maciej Zawadzki, Adam Butler

ReSolve Asset Management

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Idiosyncratic Market Effects in Global Time Series Momentum".

Warren: Here's the gist of the paper. Time series momentum models fit at the individual market level exhibit idiosyncratic forward return characteristics that are distinct from global linear and nonlinear time series momentum factor models The proposed framework enables each feature-market pair to express its unique prediction behaviour Empirical results demonstrate that fitting nonlinear models to individual markets is a robust and often overlooked modelling approach that can result in a marginal improvement in out-of-sample performance Model Framework The methodology involves casting the market return forecast as an empirical conditional estimation problem The response variable, Y = y(t)T t=1, is defined as the forward volatility-adjusted average market return The feature variables, X = x(t)T t=1, consist of time series momentum feature values The authors consider two statistical models: simple linear regression and Nadaraya-Watson kernel regression Simple linear regression models the response function as a multiple of the feature variable, yk(t) = β xk(t) + ϵ(t) Kernel regression models the response function, yk(t), as a convex combination of response values in the neighbourhood of xk(t) Experimental Design and Results The experiment data considers an asset universe of 50 global futures markets spanning over 30 years of market data Feature variables are defined as time series momentum feature values, which are winsorized at a value of ±10 and further binned into 256 equally spaced bins Response variables are defined as the forward volatility-adjusted average return, yk(t) = 1/√m mX−1 i=0 rk(t+i+l)/σˆ(t+i)k Models are compared across four specifications: Global Linear, Global Nonlinear, Sector Nonlinear, and Market Nonlinear Performance Comparison Historical walkforward simulations spanning over 30 years of market data indicated that models fit at the individual market level generalized to the out-of-sample period For the 1-month time series momentum feature, nonlinear models produced Sharpe ratios in the 0.10–0.

Title and authors: Ruth: So what does this framework actually do in practice? Essentially, they are treating the forward volatility-adjusted average market return as the response variable you want to forecast based on those momentum features.

Warren: And they’re showing that when you fit these nonlinear models at the individual market level, those response curves are distinctly different from what a single global factor model would predict.

Iris: That suggests there’s this heterogeneity in how markets respond to the momentum factor itself, which is something we need to test for identification—what specific variation is actually doing the work here?

Ruth: The paper uses historical walk-forward simulations across fifty global futures markets spanning over thirty years of data just to show this. They confirm that fitting nonlinear models at the individual market level is a robust approach, even with smaller sample sizes, and it can sometimes lead to a marginal improvement in out-of-sample performance.

Warren: The results they show for the one month time series momentum feature are pretty telling: the nonlinear models produced Sharpe ratios between zero point one zero and zero point one.

Iris: So, while the improvements are described as marginal, it’s significant because it validates the idea that we shouldn't be forced to use a single global model when markets have these distinct individual tendencies.

The paper's summary: Ruth: The authors suggest a few key ways they improve on the existing approach. They are moving beyond just simple linear regression to explore non-parametric kernel regression as a primary method for fitting those response functions.

Warren: That’s what lets them model those nonlinear response curves more accurately, which is the core improvement over sticking with just a straight line approximation.

Iris: I think the real conceptual improvement is shifting the focus from fitting one overarching global factor to allowing each feature-market pair to express its unique prediction behavior.

Ruth: Right. It’s about acknowledging that time series momentum isn't acting the same way across all markets; it changes depending on whether you’re looking at bonds, currencies, or commodities.

Warren: And the paper hints at this heterogeneity by noting that financial markets like bonds and currencies tend to show stronger mean-reversion characteristics when the momentum factor is at extreme levels compared to commodity markets.

Iris: That’s a concrete observation they make about different asset classes responding differently to the same input feature, which gives us something specific to check for identification.

The paper's improvements: Ruth: To wrap up, the main implication of this paper is that we need a more granular view of market dynamics. We shouldn't assume a single global model will capture everything because it misses these idiosyncratic effects.

Warren: So, the conclusion is that fitting nonlinear models at the individual market level is a robust way to explore these unique prediction behaviors, even if the out-of-sample gains are only marginal in some cases.

Iris: From an econometric perspective, it suggests that we should be cautious about applying generalized factor models when we know there’s a lot of variation happening at the market level.

Ruth: It’s about acknowledging that those individual market responses are important, and this paper provides a framework for doing that analysis using time series momentum features.

Warren: So, in short, they confirm that these distinct patterns exist and offer a way to model them specifically for each market.

Conclusion: Iris: Before we wrap up on "Idiosyncratic Market Effects in Global Time Series Momentum," I just want to reiterate that the work provides a concrete framework for testing the hypothesis that individual markets have distinct response curves to momentum, which is important for understanding where our models might be breaking down.

Ruth: Exactly. It’s about moving from a monolithic view of global factors to one that respects the unique quirks of each market when looking at time series momentum.

Warren: And for practical application, it suggests that if you’re building a strategy based on momentum, you should probably think about how that relationship changes depending on whether your specific asset is a bond or a commodity.

Iris: That distinction is key for any identification effort we want to make moving forward. It tells us what variation matters most.

Ruth: So, this paper gives us tools to look closer at those market-specific dynamics using kernel regression and linear regression as starting points for exploration of those individual effects.

More episodes

← Home