HTML: Hierarchical Transformer-based Multi-task Learning for Volatility Prediction

Abstract

The volatility forecasting task refers to predicting the amount of variability in the price of a financial asset over a certain period. It is an important mechanism for evaluating the risk associated with an asset and, as such, is of significant theoretical and practical importance in financial analysis. While classical approaches have framed this task as a time-series prediction one -using historical pricing as a guide to future risk forecasting -recent advances in natural language processing have seen researchers turn to complementary sources of data, such as analyst reports, social media, and even the audio data from earnings calls. This paper proposes a novel hierarchical, transformer, multi-task architecture designed to harness the text and audio data from quarterly earnings conference calls to predict future price volatility in the short and long term. This includes a comprehensive comparison to a variety of baselines, which demonstrates very significant improvements in prediction accuracy, in the range 17% -49% compared to the current state-of-the-art. In addition, we describe the results of an ablation study to evaluate the relative contributions of each component of our approach and the relative contributions of text and audio data with respect to prediction accuracy.

Multimodal Multi-Task Financial Risk Forecasting

“…Thus, making prediction tasks tough and risk-oriented [2]. Newer studies [77,94] illustrated the gains obtained by using vocal cues from the CEO's earnings conference calls for volatility prediction. Leveraging deep multimodal neural networks, they better extract the interplay between text and audio features leading to improved performance.…”

“…Recent advances in deep learning have led to a notable improvement in using multiple modalities for solving tasks such as emotion recognition [17], and audio-visual speech recognition [66]. Until recently, multimodal MTL has not been explored in the finance except for [94], which concentrates on the homogeneous tasks of stock volatility regression across various durations. We build upon their homogeneous approach to a heterogeneous multitask [95] ensemble approach with the introduction of stock price classification, owing to the relatedness of financial tasks [67].…”

“…For each stock , there exists a closing price that represents the price of at the end of the day . Formalizing stock volatility: Following [49,77,94], we define stock volatility as the natural log of the standard deviation of return prices in a window of days. This form of volatility, known as realized volatility [2,63] for the period − to , is formalized as:…”

“…2. We add more affective state-related features such as APQ (Amplitude Perturbation Quotient) Shimmer, Jitter, Spectral Density Energy, Voiced frames, Voiced to unvoiced ratio to the feature set used by works [77,94]. The resultant input vector of max( ) × where max( ) is the maximum length of conference clips and is the feature length ( = 26, max( ) = 520 audio clips, conference calls with less than maximum number of clips were padded).…”

“…Financial tasks, such as predicting a stock's price movement and volatility, are often strongly correlated, thus making multi-task learning a promising modeling choice for financial forecasting. Though homogeneous multi-task approaches [67,94] have been studied for volatility and price movement prediction, there has been no research on heterogeneous Multi-Task Learning (MTL). Here strongly related tasks on the same feature space differ based on being classification and regression problems [95,98].…”