Fig 1.
Network of the stocks for different years and their properties for the long time scale.
Fig 2.
Evolution of the network properties as a function of the year (black curves) and the stock return rescaled (blue curves).
The number of observed time scale is Nobs = 30. The vertical dashed lines correspond to (from left to right): Asian Financial Crisis, 1997 - Dotcom Crash, 2000 - Subprime Crisis, 2008 - Federal Reserve’s QE3 Announcement, 2012 - COVID-19 pandemic, 2020.
Fig 3.
Correlation matrix of the log-return of all SP stocks with the network properties on 30 observations (in years).
Table 1.
Summary of significant SSR F-test p-values (p<0.05) for Granger causality on log return with the global network variables on the long term period. The Lag represents the time delay that provides the strongest causal relationship. The ’SSR F-test p-value’ indicates the statistical significance of the causality, with lower values suggesting stronger evidence against the null hypothesis of no causal relationship.
Fig 4.
Evolution of the network properties (black curves) and the stock returns rescaled (blue curves) on short time.
The short time is two business days smoothed over a rolling window W = 23. The number of observed time scale is Nobs = 238. The vertical dashed lines correspond to (from left to right): Market Volatility Amid Inflation Concerns, Oct 2022 - SVB Bank Collapse, March 2023 - US Debt Ceiling Agreement, June 2023 - Tech Sector Rally, July 2023 - Federal Reserve Interest Rate Hike, Sept 2023 - Market Correction Due to Rising Bond Yields, Oct 2024.
Fig 5.
Correlation matrix of the log-return of all SP stocks with the network properties.
Table 2.
Best Lag and SSR F-test p-value for each global network Variable on the short time period. The Best Lag represents the time delay that provides the strongest causal relationship. The ’SSR F-test p-value’ indicates the statistical significance of the causality, with lower values suggesting stronger evidence against the null hypothesis of no causal relationship.
Table 3.
Best lag and SSR F-test p-value for each variable for the long time scale forecasting. The Best Lag represents the time delay that provides the strongest causal relationship. The ’SSR F-test p-value’ indicates the statistical significance of the causality, with lower values suggesting stronger evidence against the null hypothesis of no causal relationship.
Fig 6.
Comparison of the predictions with the log return for 5 randomly selected stocks from the testing set on the long time scale.
Differences between the wA prediction and the log return (green curves), the LRbase and the log return (orange curves) and the mean of the log return with itself (grey curves). The horizontal lines correspond to the mean of the differences.
Fig 7.
Score distribution for the predicted stock returns on long time scale.
R2 distribution (top panel) and MAE distribution (lower panel) computed using the testing set of stocks for the long time scale forecasting. Vertical lines correspond to the median values. The added value of the network parameters can be observed by comparing the base models (without network features) with the others.
Table 4.
Comparison of the Score and Mean Absolute Error (MAE) for the long time scale forecasting.
Table 5.
Best lag and SSR F-test p-value for each variable for the short time scale forecasting. The Best Lag represents the time delay that provides the strongest causal relationship. The ’SSR F-test p-value’ indicates the statistical significance of the causality, with lower values suggesting stronger evidence against the null hypothesis of no causal relationship.
Fig 8.
Score distribution for the predicted stock returns on short time scale.
R2 distribution (top panel) and MAE distribution (lower panel) computed using the testing set of stocks for the long time scale forecasting. Vertical lines correspond to the median values.The added value of the network parameters can be observed by comparing the base models (without network features) with the others.
Table 6.
Comparison of the Score and Mean Absolute Error (MAE) for the short time scale forecasting.