Manchester and Oxford's FinText Beats HAR Volatility Models by 24% on High-Risk Days
A CNN reads financial news headlines through a domain-specific word embedding, then combines with HAR benchmarks to lift realized volatility forecasts by up to 6.3%.
ProTakeaways
- Manchester and Oxford researchers train FinText, a Word2Vec/FastText embedding on 4.32B words of Dow Jones Newswires text
- A simple CNN reads news headlines and forecasts next-day realized volatility, using filter sizes 1, 2, 3 for unigram/bigram/trigram patterns
- News-only NLP underperforms CHAR benchmark, with MSE ratios of 1.106 to 1.187 across specifications
- Averaging CHAR and NLP forecasts cuts MSE by 3.9% and QLIKE by 6.3%, with realized utility rising from 2.75% to 2.93%
- Gains concentrate on high-volatility days for firm news and normal-volatility days for general news
- SHAP attribution surfaces analyst upgrades, earnings events, Fed officials, and macro indicators as top volatility drivers
Volatility forecasting has been stuck in a rut for over a decade. The workhorse HAR family models are elegant linear regressions on past realized variance that quietly assume the market's reaction to news is fully summarized by yesterday's price wiggles. A new paper from Manchester and Oxford pushes back on that assumption by feeding news headlines directly into a convolutional network, using a financial word embedding the authors trained themselves, and asking whether text carries volatility signal that HAR misses.
The short answer: on its own, text loses to HAR. Bolt the two together and you get consistent statistical and economic gains, especially on the days that matter most.
Why HAR needed help
Heterogeneous autoregressive (HAR) models parsimoniously capture volatility persistence across multiple horizons, and variants like CHAR, SHAR, HAR-J, and HARQ have piled on refinements over the years. What they don't do is look at what actually happened in the world. Media content and tone have been shown to predict market activity and price pressure, consistent with news conveying economically relevant information beyond standard market variables, yet most volatility work treats information arrival as a latent time-series phenomenon.
The paper's premise: if you can turn news text into a good numerical representation, a neural net should be able to read it and forecast next-day realized variance directly, no lagged RV required.
FinText speaks the language of markets
Off-the-shelf embeddings like Google's Word2Vec (100B tokens) and WikiNews (16B tokens) know a lot about the world but stumble on finance. The authors trained FinText on all news stories from the Dow Jones Newswires Text News Feed from 1 January 2000 to 11 September 2015, producing a corpus with 2,733,035 unique tokens and roughly 4.32 billion words. Four variants were trained: Word2Vec and FastText, each in skip-gram and CBOW flavors.