Inferensys

Glossary

Exponentially Weighted Moving Average (EWMA)

A statistical method for forecasting volatility and correlation that applies exponentially decreasing weights to historical observations, prioritizing recent data to make risk parity allocations more responsive to current market conditions.
Risk analyst performing AI risk assessment on laptop, risk matrices visible, casual office risk session.
VOLATILITY FORECASTING

What is Exponentially Weighted Moving Average (EWMA)?

A statistical method for forecasting volatility and correlation that assigns exponentially decreasing weights to historical observations, making recent data points more influential.

The Exponentially Weighted Moving Average (EWMA) is a time-series forecasting model that computes a weighted average of past squared returns, where the weights decline exponentially as observations recede into the past. Unlike a simple historical average, EWMA applies a smoothing parameter (lambda) to prioritize recent market shocks, making it highly responsive to volatility clustering and sudden regime shifts in financial markets.

In risk parity strategies, EWMA is the standard estimator for generating the ex-ante covariance matrix because it captures the dynamic, time-varying nature of asset correlations without requiring complex parameter estimation. By reacting quickly to market stress, EWMA ensures that risk contribution calculations reflect the current environment, preventing stale volatility estimates from distorting portfolio rebalancing decisions.

Volatility Forecasting

Key Characteristics of EWMA

The Exponentially Weighted Moving Average (EWMA) model is a foundational time-series method for forecasting volatility and correlation. By applying a decay factor (λ), it assigns geometrically decreasing weight to older observations, making risk parity allocations highly responsive to recent market shocks.

01

The Decay Factor (λ)

The single parameter λ (lambda) controls the persistence of the model. A value close to 1 (e.g., 0.97) creates a slow decay, retaining memory of distant events for stable long-term forecasts. A lower value (e.g., 0.90) creates a fast decay, making the model highly reactive to immediate market jumps.

  • RiskMetrics Standard: The industry benchmark sets λ = 0.94 for daily volatility and λ = 0.97 for monthly.
  • Effective Window: The half-life of a shock is calculated as ln(0.5) / ln(λ).
λ = 0.94
RiskMetrics Daily Standard
02

Recursive Calculation

EWMA avoids storing long historical windows by using a recursive formula. Today's variance estimate depends only on yesterday's variance and today's squared return. This makes it computationally lightweight for high-frequency systems.

  • Formula: σ²_t = λ * σ²_{t-1} + (1 - λ) * r²_t
  • Memory Efficiency: Only the previous period's variance needs to be cached, enabling real-time updates on streaming tick data.
O(1)
Memory Complexity
03

Reactivity to Volatility Clustering

EWMA naturally captures volatility clustering—the empirical phenomenon where large price changes tend to follow large changes. Because recent large returns immediately inflate the variance forecast, the model automatically adjusts risk budgets upward during turbulent regimes.

  • Regime Responsiveness: Risk parity weights shift rapidly to de-leverage assets with spiking EWMA volatility.
  • Limitation: EWMA does not model mean-reversion in volatility (unlike GARCH), so forecasts remain elevated until the shock decays out of the window.
04

Covariance Matrix Estimation

EWMA extends naturally to multivariate covariance for risk parity. The covariance between asset i and j is updated using the cross-product of their returns, weighted by the same decay factor.

  • Formula: σ_{ij,t} = λ * σ_{ij,t-1} + (1 - λ) * r_{i,t} * r_{j,t}
  • Consistency: Using a single λ across all pairs ensures the resulting matrix is positive semi-definite, a critical requirement for convex optimization solvers.
05

Integrated Correlation Impact

By updating the covariance matrix with EWMA, the implied correlation between assets also becomes dynamic. A sudden co-movement spike increases the covariance term, raising the correlation estimate and signaling a breakdown in diversification precisely when risk parity needs to rebalance.

  • Correlation Breakdown: During crises, correlations tend toward 1. EWMA captures this drift faster than simple historical averages.
  • Risk Budget Reallocation: Higher correlations reduce the Effective Number of Bets (ENB), prompting a reduction in leverage.
06

Lookback Window vs. Decay Factor

Unlike a simple rolling window that applies equal weight to all observations and then drops them abruptly, EWMA applies a smooth decay. This eliminates the ghosting effect where a large shock suddenly drops out of the window, causing an artificial volatility cliff.

  • Smooth Transition: The geometric decay ensures no single day causes a discontinuous jump in the risk forecast.
  • Parameter Selection: The half-life (e.g., 11 days for λ=0.94) effectively defines the 'memory' of the system without hard cutoffs.
EWMA EXPLAINED

Frequently Asked Questions

Clear, technically precise answers to the most common questions about the Exponentially Weighted Moving Average model and its role in modern risk parity and volatility forecasting.

An Exponentially Weighted Moving Average (EWMA) is a statistical forecasting model that assigns geometrically decreasing weights to historical observations, ensuring the most recent data points exert the greatest influence on the current estimate. Unlike a simple moving average that applies equal weight to all observations in a fixed window, the EWMA model applies a smoothing parameter (lambda, λ) between 0 and 1. A higher lambda gives more weight to distant history, while a lower lambda makes the model highly reactive to recent shocks. In risk management, the EWMA formula for variance is typically specified as σ²_t = λ * σ²_{t-1} + (1-λ) * r²_{t-1}, where r is the return. This recursive structure makes it computationally efficient because it only requires the previous period's variance estimate and the latest squared return, eliminating the need to store long historical time series. The model inherently captures volatility clustering—the empirical phenomenon where large market moves tend to be followed by large moves—because a spike in immediately inflates the variance forecast, which then decays exponentially as time passes without further shocks.

Prasad Kumkar

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.