10 RNN models for evaluating financial indices: examining volatility and demand-supply shifts in financial markets during COVID-19
-
Ray R. Hashemi
Abstract
Several semi-recurrent neural networks (RNNs) were developed to explore the impact of the COVID-19 pandemic on the financial indices representing bond market (yield spread - Spread), energy market (crude oil prices - Oil), stock market (volatility index - VIX and Wilshire 5,000 total market index - Wil5,000), housing market (Shiller housing price index - HPI), commodity market (Gold, Wheat, and Soybean indices), and recession (USrec). The indices were divided into two groups based on two properties: volatility and demand-supply shifts. USrec was a part of both groups. Two time-series datasets (dataset 1 and dataset 2) were obtained from the Federal Reserve Bank of St. Louis for the two groups. The volatility and demand-supply shift groups had daily and monthly data, respectively. Each dataset was first purified by applying derived-breakpoints (the transition block) for each index, separately. Second, the purified dataset was partitioned into “before” and “during” the pandemic. Semi-RNNs were trained by the training sets for the “before” and “during”, separately, and tested against their corresponding test sets which delivered two 5-element counter arrays. Each array carries the overall percent of the correct classifications of the indices for the entire records of the corresponding test sets. A decay rate was calculated for each index in the arrays and the average of decay rates (overall decay) is also calculated for each array to be used for inspecting the impact of COVID-19 on indices. The positive, zero, and negative decay rate means COVID-19 has negative, none, and positive impacts on the financial indices, respectively. The results also revealed: (a) VIX was explained by Wil5,000, Spread, Oil, and USrec more accurately before the pandemic, indicating that other observed and unobserved factors arising from the COVID-19 pandemic would affect the VIX more than the other financial indices. (b) USrec is predicted less accurately compared to other indices during the pandemic, which shows the sensitivity of this index to health and geopolitical challenges. (c) Effects of the indices on the bond market diminished during the pandemic. (d) HPI declines in predic tive accuracy during the pandemic indicate disruptions in the housing market due to economic and social changes. (e) Gold price drop in predictive accuracy during the pandemic reflects its nature as a relatively stable safe-haven but still sensitive to COVID-19. (f) Wheat’s substantial drop in predictive accuracy highlights its sensitivity to disrupted supply chains, global trade dynamics, and fluctuating demands during the pandemic. (g) Similar to wheat, soybean predictive accuracy fell during the pandemic, underscoring the impact of supply chain disruptions on agricultural commodities. The sensitivity analysis suggests that the results remain highly robust to the changes in the number of records chosen for the different test sets.
Abstract
Several semi-recurrent neural networks (RNNs) were developed to explore the impact of the COVID-19 pandemic on the financial indices representing bond market (yield spread - Spread), energy market (crude oil prices - Oil), stock market (volatility index - VIX and Wilshire 5,000 total market index - Wil5,000), housing market (Shiller housing price index - HPI), commodity market (Gold, Wheat, and Soybean indices), and recession (USrec). The indices were divided into two groups based on two properties: volatility and demand-supply shifts. USrec was a part of both groups. Two time-series datasets (dataset 1 and dataset 2) were obtained from the Federal Reserve Bank of St. Louis for the two groups. The volatility and demand-supply shift groups had daily and monthly data, respectively. Each dataset was first purified by applying derived-breakpoints (the transition block) for each index, separately. Second, the purified dataset was partitioned into “before” and “during” the pandemic. Semi-RNNs were trained by the training sets for the “before” and “during”, separately, and tested against their corresponding test sets which delivered two 5-element counter arrays. Each array carries the overall percent of the correct classifications of the indices for the entire records of the corresponding test sets. A decay rate was calculated for each index in the arrays and the average of decay rates (overall decay) is also calculated for each array to be used for inspecting the impact of COVID-19 on indices. The positive, zero, and negative decay rate means COVID-19 has negative, none, and positive impacts on the financial indices, respectively. The results also revealed: (a) VIX was explained by Wil5,000, Spread, Oil, and USrec more accurately before the pandemic, indicating that other observed and unobserved factors arising from the COVID-19 pandemic would affect the VIX more than the other financial indices. (b) USrec is predicted less accurately compared to other indices during the pandemic, which shows the sensitivity of this index to health and geopolitical challenges. (c) Effects of the indices on the bond market diminished during the pandemic. (d) HPI declines in predic tive accuracy during the pandemic indicate disruptions in the housing market due to economic and social changes. (e) Gold price drop in predictive accuracy during the pandemic reflects its nature as a relatively stable safe-haven but still sensitive to COVID-19. (f) Wheat’s substantial drop in predictive accuracy highlights its sensitivity to disrupted supply chains, global trade dynamics, and fluctuating demands during the pandemic. (g) Similar to wheat, soybean predictive accuracy fell during the pandemic, underscoring the impact of supply chain disruptions on agricultural commodities. The sensitivity analysis suggests that the results remain highly robust to the changes in the number of records chosen for the different test sets.
Kapitel in diesem Buch
- Frontmatter I
- Preface V
- Contents VII
-
Methods and instrumentation
- 1 Identifying and estimating outliers in time series with nonstationary mean through multiobjective optimization method 1
- 2 Using the intentionally linked entities (ILE) database system to create hypergraph databases with fast and reliable relationship linking, with example applications 21
- 3 Rapid and automated determination of cluster numbers for high-dimensional big data: a comprehensive update 37
- 4 Canonical correlation analysis and exploratory factor analysis of the four major centrality metrics 49
- 5 Navigating the landscape of automated data preprocessing: an in-depth review of automated machine learning platforms 71
- 6 Generating random XML 83
-
Applications and case studies
- 7 Exploring autism risk: a deep dive into graph neural networks and gene interaction data 105
- 8 Leveraging ChatGPT and table arrangement techniques in advanced newspaper content analysis for stock insights 121
- 9 An experimental study on road surface classification 145
- 10 RNN models for evaluating financial indices: examining volatility and demand-supply shifts in financial markets during COVID-19 165
- 11 Topological methods for vibration feature extraction 185
- 12 Dyna-SPECTS: DYNAmic enSemble of Price Elasticity Computation models using Thompson Sampling in e-commerce 215
- 13 Creating a metadata schema for reservoirs of data: a systems engineering approach 251
- 14 Implementation and evaluation of an eXplainable artificial intelligence to explain the evaluation of an assessment analytics algorithm for freetext exams in psychology courses in higher education to attest QBLM-based competencies 271
- 15 Toward a skill-centered qualification ontology supporting data mining of human resources in knowledge-based enterprise process representations 307
- Index 333
Kapitel in diesem Buch
- Frontmatter I
- Preface V
- Contents VII
-
Methods and instrumentation
- 1 Identifying and estimating outliers in time series with nonstationary mean through multiobjective optimization method 1
- 2 Using the intentionally linked entities (ILE) database system to create hypergraph databases with fast and reliable relationship linking, with example applications 21
- 3 Rapid and automated determination of cluster numbers for high-dimensional big data: a comprehensive update 37
- 4 Canonical correlation analysis and exploratory factor analysis of the four major centrality metrics 49
- 5 Navigating the landscape of automated data preprocessing: an in-depth review of automated machine learning platforms 71
- 6 Generating random XML 83
-
Applications and case studies
- 7 Exploring autism risk: a deep dive into graph neural networks and gene interaction data 105
- 8 Leveraging ChatGPT and table arrangement techniques in advanced newspaper content analysis for stock insights 121
- 9 An experimental study on road surface classification 145
- 10 RNN models for evaluating financial indices: examining volatility and demand-supply shifts in financial markets during COVID-19 165
- 11 Topological methods for vibration feature extraction 185
- 12 Dyna-SPECTS: DYNAmic enSemble of Price Elasticity Computation models using Thompson Sampling in e-commerce 215
- 13 Creating a metadata schema for reservoirs of data: a systems engineering approach 251
- 14 Implementation and evaluation of an eXplainable artificial intelligence to explain the evaluation of an assessment analytics algorithm for freetext exams in psychology courses in higher education to attest QBLM-based competencies 271
- 15 Toward a skill-centered qualification ontology supporting data mining of human resources in knowledge-based enterprise process representations 307
- Index 333
