Discussing the article: "MQL5 Wizard Techniques you should know (14): Multi Objective Timeseries Forecasting with STF"
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Check out the new article: MQL5 Wizard Techniques you should know (14): Multi Objective Timeseries Forecasting with STF.
Spatial Temporal Fusion which is using both ‘space’ and time metrics in modelling data is primarily useful in remote-sensing, and a host of other visual based activities in gaining a better understanding of our surroundings. Thanks to a published paper, we take a novel approach in using it by examining its potential to traders.
This paper on Spatial Temporal Fusion (STF) piqued my interest on the subject thanks to its two-sided approach to forecasting. For a refresher, the paper is inspired by solving a probability-based forecasting problem that is collaborative for both supply and demand in two-sided ride-hailing platforms, such as Uber and Didi. Collaborative supply and demand relationships are common in various two-sided markets, such as Amazon, Airbnb, and eBay where in essence the company not only serves the traditional ‘customer’ or purchaser, but also caters to suppliers of the customer.
So, two-sided forecasting in a case where supply is partly dependent on demand can be important to these companies on a frequent basis. This dual projection though, of demand and supply, was certainly a break from the conventional approach of forecasting a specific value to a timeseries or data set. The paper also introduced what it called a causaltrans framework where the causal ‘collaborative’ relationship between supply and demand was captured by a matrix G and all forecasts were made via transformer network and its results were noteworthy.
Taking a leaf from that, we look to forecast supply and demand for traded securities by using bearishness and bullishness as proxies for these two metrics. Strictly speaking though the typical Expert-Signal class does compute both these values as integers in the range 0-100 as can be seen in the MQL5 library files or files we have coded in these series so far. What would be new though will be the addition of a spatial matrix and a time parameter in making our forecasts (the 2 extra inputs we cite from the paper).
Spatial quantization of trade securities is subjective and so is the choice of time metric. Using security High-prices and Low-price series as our anchors for demand and supply, we use the autocorrelation values amongst these buffers as coordinates to a spatial matrix as well as the day of week index as a time indicator. This rudimentary approach that can be customized and improved, serves our purposes for this article.
The paper used transformer networks which we will not use as it is inefficient for our purposes however all forecasts will be through a custom hand coded multi-layer perceptron. With so many libraries and code samples on the subject, it would seem a waste of time to attempt to code one’s own multilayer perceptron. However the network class used is less than 300 lines long and is reasonably scalable in as far as customizing the number of layers and size of each, something which is still lacking in most of the boiler plate libraries that are available.
Author: Stephen Njuki