Robust estimation of statistical temporal networks for financial time series modeling: theoretical formulation and temporal patterns typology

Although the idea of using Neural Networks technology for Financial Time Series prediction is an old one, the abundance and availability of stock-price data, including high-frequency (intra-minute) data has given an additional impetus to this fledgling field of study. Nevertheless, the study and applications have focused on the hedge-funds and brokerage operations of investments banks and very little academic attention has been devoted to the day-trading activity aiming at the accumulation of short-term incremental gains – still considered as a retail activity, similar to betting. The purpose of this research is precisely to investigate the possibility to use the sound time series smoothing techniques of feedforward Neural Networks along with elementary but powerful Classification Neural Networks techniques to produce a decision-aid system for intraday trading decisions. The basic design of the study consisted in presenting a new typology of intraday patterns and apply it to the 126 trading days of the first half of 2020 for the SP500 index, using intraday, minute-by-minute prices. While applying this methodology generated questions for further research, the major finding of this study was that when a price trend was soundly established during the first half of a trading day session, more often than not (in 57 cases versus 31) the trend continued till the end of the trading day. This result, which translates into the possibility for the trader to engage in short-term profitable trades, is non-trivial, though needs more past data to be consolidated. Further research into the conditions prevailing in other markets, before a trading day begins, could also prove useful.