Hybrid ARDL-MIDAS-Transformer time-series regressions for multi-topic crypto market sentiment driven by price and technology factors

Abstract This paper develops a novel hybrid Autoregressive Distributed Lag Mixed Data Sampling (ARDL-MIDAS) model that integrates both deep neural network multi-head attention Transformer mechanisms, and number of covariates, including sophisticated stochastic text time-series features, into mixed-frequency regression with long memory structure. In doing so, we demonstrate how the resulting class ARDL-MIDAS-Transformer models allows one to maintain interpretability whilst exploiting architectures. The latter may be used for higher-order interaction analysis, or, as in our use case, design Instrumental Variables reduce bias estimation infinite lag ARDL-MIDAS model. Our approach produces an accurate, interpretable forecasting framework forecast end-of-day sentiment intra-daily, readily attainable regressors. this regard, conduct statistical analysis on mixed data frequencies discover study relationships between from custom framework, alternative popular extraction frameworks (BERT VADER), technology factors, well investigate role price discovery has retail cryptocurrency investors’ (crypto sentiment). is interesting modelling challenge it involves working which response process, are observed at different time scales. Specifically, detailed real-data conducted where explore relationship daily crypto market (of positive, negative neutral polarity) intra-daily (hourly) log-return dynamics markets. indices constructed variety “topics” news sources produced collection capturing polarity signals each “topic”, namely particular or asset. Different methods developed context, utilised proposed framework. Furthermore, factors introduced capture effects, such hash rate important aspect money supply relating mining new assets, block hashing transaction verification. Throughout real study, provide guidance insights combine—in transparent, non-black-box way—covariates obtained resolutions, understand arising these potentially under presence structure, and, finally, successfully leverage applications. demonstrated superior performance alternatives in-sample application data.