Forecasting Volatility with Nontraditional Data

Overview

This project explores whether alternative data sources can improve volatility forecasting beyond traditional GARCH models.

Methodology

We combine classical volatility models with features derived from:

Economic news sentiment
Options market implied volatility
Social media activity metrics

R: Enhanced GARCH Model▼

library(rugarch)
library(quantmod)
library(tidyverse)

Standard GARCH(1,1) specification
spec_base <- ugarchspec(
  variance.model = list(model = "sGARCH", garchOrder = c(1, 1)),
  mean.model = list(armaOrder = c(1, 0)),
  distribution.model = "std"
)
GARCH-X with external regressors
spec_enhanced <- ugarchspec(
  variance.model = list(
    model = "sGARCH", 
    garchOrder = c(1, 1),
    external.regressors = as.matrix(alt_data[, c("vix_lag", "news_sentiment", "social_volume")])
  ),
  mean.model = list(armaOrder = c(1, 0)),
  distribution.model = "std"
)
Fit models
fit_base <- ugarchfit(spec_base, returns)
fit_enhanced <- ugarchfit(spec_enhanced, returns)
Compare information criteria
cat("Base AIC:", infocriteria(fit_base)[1], "\n")
cat("Enhanced AIC:", infocriteria(fit_enhanced)[1], "\n")

Results

The enhanced model with external regressors achieved:

8% lower RMSE in out-of-sample volatility forecasts
Particularly strong improvement during market stress periods