A mixed-frequency approach for exchange rates predictions

This paper proposes a mixed-frequency modeling approach to overcome the information loss inherent in temporal aggregation, demonstrating its effectiveness in predicting CAD/USD exchange rates and addressing the Meese and Rogoff puzzle.

The Gist

Imagine you are trying to predict the weather for next month.

The Old Way (The Puzzle):
For decades, economists have been trying to predict exchange rates (the price of one country's money compared to another, like the Canadian Dollar vs. the US Dollar). They have used fancy theories about interest rates, inflation, and money supply. But there's a famous problem called the "Meese and Rogoff Puzzle." It's like a stubborn riddle that says: "No matter how hard you try with your complex theories, a simple guess that 'tomorrow will look exactly like today' (a Random Walk) is actually just as good, or even better, than your fancy models."

Why is this happening? The authors of this paper suggest the problem isn't the theories; it's the magnifying glass they are using.

The Problem: Blurring the Picture
Most studies look at economic data once a month or once a quarter (every three months). But the real world moves much faster. Interest rates and prices change every day.

Think of it like watching a high-speed race car on TV.

• The Real World: The car is zooming by at 200 mph. You can see every bump, every turn, and every gear shift.
• The Old Studies: They take a photo of the car only once every hour. When they try to predict where the car will be next, they are working with a blurry, low-resolution image. They've thrown away all the exciting details that happened in between the photos. This is called "Temporal Aggregation Bias." They are trying to solve a high-speed puzzle with a low-speed map.

The New Solution: The Mixed-Frequency Approach (MIDAS)
The authors, Mattera, Misuraca, and their team, decided to stop squinting at the blurry photos. They used a new statistical tool called MIDAS (Mixed Data Sampling).

Think of MIDAS as a super-lens that lets you see the race car in high definition while still predicting where it will be an hour from now.

• They kept the "destination" (the exchange rate) as a quarterly check-in.
• But instead of ignoring the daily and monthly speed bumps, they fed all that high-speed data (daily interest rates, monthly inflation) into the model.

It's like having a GPS that doesn't just know the road map, but also knows the traffic jams, the construction zones, and the weather conditions happening right now, even if you only check your destination once an hour.

The Experiment: The CAD/USD Race
They tested this on the Canadian Dollar vs. the US Dollar.

1. The Control Group: They ran the old models (using only quarterly data). As expected, the "Random Walk" (the simple guess) won. The complex models failed.
2. The Test Group: They ran the new MIDAS models (using all the daily/monthly data).

The Result:
The new models crushed the competition.

• When they used the "super-lens" (MIDAS), the models based on interest rates and inflation suddenly became very accurate.
• Specifically, the "Taylor Rule" model (which looks at how central banks adjust interest rates based on inflation) became 18% more accurate with the new method.
• The "Sticky Price" model (a complex theory about how prices adjust slowly) became a massive 53% more accurate.

The Takeaway
The paper concludes that the "Meese and Rogoff Puzzle" wasn't a mystery about the economy being unpredictable. It was a mystery about bad data handling.

The economy is predictable, but only if you stop throwing away the fast-moving details. By using a mixed-frequency approach, the authors showed that if you listen to the high-speed chatter of the daily market, you can actually predict the future of exchange rates much better than the simple "guess it stays the same" method.

In a nutshell: They fixed the blurry camera, and suddenly, the future became much clearer.

Technical Summary

Here is a detailed technical summary of the paper "A mixed-frequency approach for exchange rates predictions" by Mattera et al. (2021).

1. Problem Statement

The paper addresses the Meese and Rogoff puzzle, a long-standing empirical anomaly in economics stating that standard economic models (based on fundamentals like interest rates, money supply, and price levels) fail to forecast exchange rates more accurately than a simple random walk model.

The authors hypothesize that the failure of these models is partly due to temporal aggregation bias. Most forecasting studies aggregate high-frequency data (daily or monthly) into lower frequencies (monthly or quarterly) to match the frequency of macroeconomic predictors. This aggregation process discards significant information contained in the high-frequency variations of the predictors. The paper argues that by forcing all variables into a single low-frequency framework, researchers omit crucial high-frequency dynamics, leading to model misspecification and poor out-of-sample performance.

2. Methodology

To overcome the limitations of temporal aggregation, the authors propose a Mixed Data Sampling (MIDAS) regression approach.

• Core Concept: MIDAS allows for the estimation of a regression model where the dependent variable (exchange rate) is observed at a low frequency (e.g., quarterly), while the independent variables (predictors) are observed at a higher frequency (e.g., monthly).
• Unrestricted MIDAS (U-MIDAS): The authors utilize the U-MIDAS framework (Foroni et al., 2015). Instead of using complex distributed lag polynomials (like Almon lags) to parameterize the relationship, U-MIDAS employs a temporal alignment strategy.
  • High-frequency data vectors are "stacked" to align with the low-frequency observation periods.
  • For a quarterly dependent variable ($y_t$) and monthly predictors ($x_t$), the model includes the current month and the two preceding months of the quarter as separate regressors.
  • This transforms the mixed-frequency problem into a standard linear regression that can be estimated via Ordinary Least Squares (OLS).
• Models Tested: The authors applied this mixed-frequency approach to several classical exchange rate theories, creating "Mixed Frequency" (MF) versions of:
  1. Uncovered Interest Rate Parity (UIRP)
  2. Purchasing Power Parity (PPP)
  3. Monetary Models (Flexible and Sticky price versions)
  4. Taylor Rule Fundamentals (Instantaneous and Smoothing versions)
• Data: The study uses CAD/USD exchange rate data from January 1985 to January 2019.
  • Dependent Variable: Quarterly exchange rate returns.
  • Predictors: Monthly data for interest rates, CPI (price levels), M3 money supply, GDP, output gaps, and inflation rates (sourced from FRED and OECD).
• Evaluation:
  • Out-of-Sample Analysis: Performed using both recursive (expanding window) and rolling-window schemes.
  • Metrics: Mean Square Forecast Error (MSFE) for accuracy.
  • Statistical Tests: Diebold-Mariano (DM) and Clark-West (CW) tests to determine if forecast improvements are statistically significant compared to the benchmark.

3. Key Contributions

1. Novel Application: This is one of the first studies to apply MIDAS regression specifically to solve the Meese and Rogoff puzzle by utilizing monthly macroeconomic fundamentals to forecast quarterly exchange rates.
2. Resolution of Aggregation Bias: The paper provides empirical evidence that the "unpredictability" of exchange rates is partly an artifact of temporal aggregation. By retaining high-frequency information, the models capture dynamics that aggregated models miss.
3. Taylor Rule Superiority: The study reinforces the importance of Taylor rule fundamentals (inflation and output gaps) in exchange rate forecasting, showing they perform robustly when combined with mixed-frequency techniques.

4. Empirical Results

The study compares classical models (aggregated to quarterly) against their Mixed-Frequency (MF) counterparts.

A. Recursive Approach (Expanding Window):

• Overall Performance: Most MF models outperformed their classical counterparts and the Random Walk benchmark.
• Best Performer: The MF-TYLR2 (Mixed Frequency Taylor Rule with smoothing) provided the most accurate forecasts among MF models.
• Significant Gains:
  • MF-MM2 (Sticky Price Monetary Model) showed a massive 53.6% improvement in accuracy over the classical MM2.
  • MF-TYLR2 was 18% more accurate than the classical Taylor Rule.
  • MF-PPP and MF-UIRP also showed significant improvements (8.2% and 2.3%, respectively).
• Statistical Significance: The Diebold-Mariano and Clark-West tests confirmed that these improvements were statistically significant.

B. Rolling Window Approach:

• The Puzzle Reappears: In the rolling window scheme, classical models (UIRP, Monetary Models) largely failed to beat the Random Walk, confirming the Meese and Rogoff puzzle in this specific setting.
• MF Success: However, the MF-UIRP and MF-PPP models significantly outperformed both the Random Walk and their classical aggregated versions.
  • MF-UIRP achieved a 10% accuracy gain over the classical UIRP, effectively explaining the unpredictability of the classical model as a result of temporal aggregation.
  • MF-TYLR2 remained the strongest performer, significantly beating the benchmark.

5. Significance and Conclusion

The paper concludes that the Meese and Rogoff puzzle is not necessarily a failure of economic theory, but rather a failure of statistical methodology regarding data frequency.

• Implication for Policymakers: Central banks and policymakers should consider utilizing mixed-frequency models to extract more accurate signals from high-frequency economic data, rather than relying solely on aggregated quarterly or monthly data.
• Theoretical Contribution: The study suggests that the "unpredictability" of exchange rates is partly a statistical artifact caused by omitting high-frequency information. By using MIDAS to incorporate all available monthly data into quarterly forecasts, the predictive power of economic fundamentals is restored.
• Practical Application: The authors demonstrate that the MIDAS framework is a viable and superior tool for exchange rate forecasting, offering a new domain for its application in international finance.