Modeling and Forecasting Exchange Rate Volatility in Time-Frequency Domain

Table of Contents

1. Introduction & Overview

This paper presents a novel approach to modeling and forecasting financial volatility, specifically for exchange rates, by integrating high-frequency data analysis with time-frequency decomposition techniques. The core innovation lies in augmenting the Realized GARCH framework with wavelet-decomposed realized volatility measures and a specialized jump estimator. This allows the model to dissect volatility into components corresponding to different investment horizons (timescales) and separately account for the impact of discontinuous price jumps. The research is motivated by the heterogeneous nature of market participants who operate on varying time horizons, from high-frequency traders to long-term investors.

The authors demonstrate that their proposed "Jump-GARCH" models, estimated via both Maximum Likelihood and the Generalized Autoregressive Score (GAS) framework, provide statistically superior forecasts compared to conventional GARCH and popular realized volatility models. The analysis uses foreign exchange futures data encompassing the 2007-2008 financial crisis, providing a robust stress test for the methodology.

2. Methodology & Technical Framework

2.1 Realized GARCH Framework

The Realized GARCH model bridges the gap between traditional GARCH models and high-frequency data by incorporating a realized volatility measure $RV_t$ directly into the volatility equation. The basic structure involves a return equation, a GARCH equation for latent volatility, and a measurement equation linking latent volatility to the realized measure.

2.2 Wavelet-Based Multiscale Decomposition

To capture the multi-horizon nature of volatility, the authors employ a wavelet transform. This mathematical tool decomposes the realized volatility series into orthogonal components representing different timescales (e.g., intraday, daily, weekly dynamics). If $RV_t$ is the realized volatility, its wavelet decomposition can be represented as:

$RV_t = \sum_{j=1}^J D_{j,t} + S_{J,t}$

where $D_{j,t}$ represents the volatility component ("detail") at scale $j$ (corresponding to a specific frequency band), and $S_{J,t}$ is the smooth component capturing the longest-term trend. Each $D_{j,t}$ approximates the trading activity and information flow at a specific investment horizon.

2.3 Jump Detection & JTSRV Estimator

A critical advancement is the integration of jump variation. The authors utilize a Jump Two Scale Realized Volatility (JTSRV) estimator. This estimator separates the total quadratic variation into the continuous integrated variance (IV) and the discontinuous jump variance (JV):

$RV_t \approx IV_t + JV_t$

This separation is crucial as jumps and continuous volatility often have different persistence and forecasting properties.

2.4 Estimation: MLE vs. GAS

The proposed Jump-GARCH models are estimated using two methods: 1) Quasi-Maximum Likelihood Estimation (QMLE), and 2) the observation-driven Generalized Autoregressive Score (GAS) framework. The GAS framework, introduced by Creal et al. (2013), updates parameters based on the score of the likelihood function, offering potential robustness and adaptability to model misspecification.

3. Empirical Analysis & Results

3.1 Data & Experimental Setup

The study uses high-frequency data for FX futures (likely major pairs like EUR/USD). The sample period includes the 2007-2009 financial crisis, allowing examination of model performance under extreme stress. Forecasts are evaluated for both one-day-ahead and multi-period-ahead horizons.

3.2 Forecasting Performance

The proposed models are benchmarked against standard models like GARCH(1,1) and HAR-RV. The evaluation uses statistical loss functions (e.g., MSE, QLIKE). The key results are presented in a comparative table (simulated below):

Note: Values are illustrative ratios relative to GARCH(1,1) benchmark.

3.3 Key Findings & Insights

• Jump Separation is Key: Disentangling jump variation from integrated variance consistently improves forecasting accuracy.
• High-Frequency Dominance: The most informative timescale for future volatility is the high-frequency (short-horizon) component of the wavelet decomposition.
• Model Superiority: The newly proposed Jump-GARCH models with wavelet decomposition statistically outperform both conventional GARCH and standard Realized GARCH models.
• Crisis Resilience: The models demonstrate robust performance during the financial crisis period.

4. Core Insight & Analyst Perspective

Core Insight: This paper delivers a powerful, yet under-appreciated, message: volatility is not a monolithic process but a layered one. By refusing to treat the market as a single, homogeneous entity and instead using wavelets to dissect it into its constituent investment horizons, the authors crack open the black box of volatility dynamics. The finding that short-term, high-frequency components drive forecasts is a direct challenge to models that overweight longer-term trends and underscores the increasing dominance of algorithmic and high-frequency trading in price discovery and volatility formation.

Logical Flow: The argument is elegantly constructed. It starts from the well-established empirical fact of heterogeneous market agents (from Corsi's HAR model). It then logically asks: if agents operate on different timescales, shouldn't our models reflect that? The wavelet decomposition is the perfect technical answer. The subsequent integration of jump risk—another non-Gaussian, discontinuous reality of markets—completes the picture. The flow from economic intuition (heterogeneity) to mathematical tool (wavelets) to empirical result (forecast improvement) is compelling.

Strengths & Flaws: The primary strength is the successful fusion of sophisticated econometrics (Realized GARCH, wavelets, jump detection) into a coherent, empirically successful framework. It moves beyond simple model comparisons to provide genuine insight into the source of predictability. The use of the GAS framework is also forward-looking. The main flaw, common in this literature, is the "in-sample" feel of the robustness check. While the crisis period is included, a true out-of-sample test on entirely unseen data (e.g., the 2020 COVID crash) would be more convincing. Furthermore, the computational complexity of the wavelet-GARCH-jump model may limit its real-time application in some trading systems, a practical hurdle not addressed.

Actionable Insights: For quants and risk managers, this paper is a blueprint. First, decompose, then model. Applying a simple wavelet filter to your volatility series before feeding it into your favorite ML or econometric model could yield immediate gains. Second, treat jumps separately. Building a dedicated signal for jump detection and modeling its impact independently, as done with the JTSRV, is a non-negotiable best practice for any serious volatility model post-2008. Finally, focus your forecasting energy on the high-frequency layer. Allocate more research and computational resources to understanding and predicting the intraday volatility dynamics, as this is where the most significant predictive signal lies.

5. Technical Details & Mathematical Formulation

The core Jump-GARCH model with wavelet components can be summarized as follows:

Return Equation: $r_t = \sqrt{h_t} z_t$, where $z_t \sim i.i.d.(0,1)$.

GARCH Equation: $h_t = \omega + \beta h_{t-1} + \gamma \xi_{t-1}$.

Measurement Equation (Enhanced):
$\log(RV_t) = \xi + \phi \log(h_t) + \tau_1 z_t + \tau_2 (z_t^2 - 1) + \sum_{j=1}^J \delta_j D_{j,t} + \lambda J_t + u_t$
where $u_t \sim i.i.d.(0, \sigma_u^2)$. Here, $D_{j,t}$ are the wavelet-detail components of $RV_t$, and $J_t$ is the significant jump component identified by the JTSRV estimator.

The model estimates the parameters $\theta = (\omega, \beta, \gamma, \xi, \phi, \tau_1, \tau_2, \{\delta_j\}, \lambda)$ to capture the dynamics between latent volatility, realized measures, jumps, and multi-scale components.

6. Analysis Framework: Example Case

Scenario: A quantitative hedge fund wants to improve its daily Value-at-Risk (VaR) forecast for a EUR/USD trading book.

Step 1 - Data Preparation: Acquire 5-minute intraday returns for EUR/USD. Calculate a baseline realized volatility (e.g., RV) and apply a wavelet transform (using a library like PyWavelets in Python) to decompose it into 3 scales: D1 (2-4 hour dynamics), D2 (4-8 hour), D3 (8-16 hour). Separately, apply the JTSRV estimator to extract the daily jump series $J_t$.

Step 2 - Model Specification & Estimation: Estimate the Jump-GARCH model from Section 5, where the measurement equation includes D1, D2, D3, and $J_t$ as exogenous variables. Compare the log-likelihood and information criteria with a standard Realized GARCH model.

Step 3 - Forecasting & Application: Generate the one-day-ahead volatility forecast $\hat{h}_{t+1}$ from the estimated model. Use this forecast to calculate the VaR (e.g., $VaR_{t+1}^{\alpha} = -\Phi^{-1}(\alpha) \sqrt{\hat{h}_{t+1}}$). Backtest the VaR forecasts against actual P&L to assess coverage accuracy.

Expected Outcome: The VaR forecasts from the Jump-GARCH model with wavelets should exhibit more accurate coverage (fewer exceptions) and be less prone to under-estimating risk following days with high jumps or specific intraday volatility patterns.

7. Future Applications & Research Directions

• Machine Learning Integration: The wavelet components $D_{j,t}$ and jump series $J_t$ could serve as highly informative features for machine learning models (e.g., LSTM, Gradient Boosting) for volatility forecasting, moving beyond the linear/parametric GARCH structure.
• Cross-Asset Volatility Spillovers: Apply the multiscale decomposition to study how volatility transmits between asset classes (e.g., from equities to FX) at different time horizons. Does a stock market crash transmit via short-term or long-term volatility components?
• Real-Time Trading Signals: Develop trading strategies that explicitly use the discrepancy between short-horizon and long-horizon volatility components as a mean-reversion or momentum signal.
• Central Bank & Policy Analysis: Use the framework to analyze the impact of monetary policy announcements on FX volatility, distinguishing between the immediate high-frequency "news spike" and the longer-term assimilation of information.
• Extension to Cryptocurrencies: Test the model on 24/7 cryptocurrency markets, which are characterized by extreme jumps and multi-scale investor behavior, from algorithmic bots to long-term "HODLers."

8. References

1. Barunik, J., Krehlik, T., & Vacha, L. (2015). Modeling and forecasting exchange rate volatility in time-frequency domain. Preprint, arXiv:1204.1452v4.
2. Corsi, F. (2009). A simple approximate long-memory model of realized volatility. Journal of Financial Econometrics, 7(2), 174-196.
3. Hansen, P. R., & Lunde, A. (2005). A forecast comparison of volatility models: does anything beat a GARCH(1,1)? Journal of Applied Econometrics, 20(7), 873-889.
4. Creal, D., Koopman, S. J., & Lucas, A. (2013). Generalized autoregressive score models with applications. Journal of Applied Econometrics, 28(5), 777-795.
5. Gençay, R., Selçuk, F., & Whitcher, B. (2005). Multiscale systematic risk. Journal of International Money and Finance, 24(1), 55-70.
6. McAleer, M., & Medeiros, M. C. (2008). A multiple regime smooth transition heterogeneous autoregressive model for long memory and asymmetries. Journal of Econometrics, 147(1), 104-119.
7. Andersen, T. G., & Bollerslev, T. (1998). Answering the skeptics: Yes, standard volatility models do provide accurate forecasts. International Economic Review, 39(4), 885-905.