Discover how Machine Learning is used by traders to invest, arbitrage, hedge or speculate on financial markets.
Machine learning, an artificial intelligence technology for traders
Traders today are using new technologies to boost their performance, especially through Machine Learning techniques.
When you choose a piece of music on Spotify or your next series on Netflix, your streaming platform uses artificial intelligence to discover what your preferences are and make the best possible recommendations. To do this, these search engines rely on mountains of data as well as innovative technology: machine learning.
But digital giants are not the only ones to use Machine learning technologies, financial professionals are also very interested in it! Let’s discover together how Machine Learning helps traders in the trading room to make decisions.
What is Machine learning?
By definition, Machine learning is an artificial intelligence technology that allows algorithms to sort through data and extract knowledge to make informed decisions.
This new technology is based on code and advanced mathematics,
but remains in fine a form of automatism aiming at recognizing certain specific situations.
Although the concept of Machine learning is often used interchangeably with the concept of Deep learning, it is important to understand the difference between these two terms.
For example, while Machine learning technologies need the user to know if their decisions are right or wrong and to learn from them, Deep learning draws its own conclusions thanks to a specific computer network (called “neural network”).
If these two fields both belong to the world of artificial intelligence and remain similar in many ways, their difference in terms of learning capabilities remains crucial.
NB: Although Machine learning has experienced a new boom in recent years with the arrival of the “Big Data” era, this technology is not new, since the first Machine learning program (the Perceptron) was developed as early as 1957!
The different Machine learning programs
Machine learning programs can be divided into two main categories:
- supervised Machine learning programs (using training data that has already been labeled during the learning process);
- unsupervised Machine learning programs (using training data that has not been labeled beforehand).
Furthermore, it is also possible to distinguish between classification algorithms (non-numerical) and regression algorithms (numerical), as well as algorithms that use or do not use reinforced learning, the mechanism of which consists of guiding the algorithm by giving it a reward when it gets closer to the objective, and conversely, imposing a penalty when it moves away from it.
Machine learning applied to trading
In the context of trading, a Machine Learning program can assist the trader in performing all kinds of automated tasks or any predictive exercise.
Unlike a classical trading algorithm that follows a precise sequence of instructions in a systematic way, a Machine Learning system will not blindly follow a sequence of instructions but will learn from the experience it is confronted with, and more precisely from the feedback it receives.
In a very concrete way, the Machine Learning technology can thus analyze the behavior of a trader to learn which market situations he prefers to trade in order to present him at the beginning of each session with the trading opportunities that best suit his tastes.
By extension, it is also possible to train a Machine Learning program to analyze not the preferences of a Trader but the market data in order to train him to recognize certain opportunities that he could take advantage of autonomously.
Whether it’s analyzing price movements to recognize specific chartist patterns, technical indicators to identify specific signals, or investor behavior to spot certain characteristic movements, Machine Learning is a formidable weapon.
And for good reason, since Machine Learning needs large amounts of data to be effective, it flourishes in an environment such as the financial markets because of the abundance of historical information available. Its ability to extract value from a data set without human intervention can therefore be fully exploited.
However, due to the amount of data required and the cost of the hardware involved, machine learning programs may primarily benefit large market players who own a considerable amount of data and are able to invest millions in developing and perfecting these solutions.
Furthermore, it is quite possible to envisage the use of Machine Learning by stock market regulators in order to detect in an automated way any suspicious behavior on the markets, such as insider trading for example.
Finally, Machine learning programs can also be used to prepare new traders by identifying the best way to train a person to make him operational on the financial markets, and able to perform such or such types of operations.
This last application may seem trivial, but it is of great interest to the world’s largest trading rooms, which are always on the lookout for the best talent to succeed in one of the most competitive professional sectors.
Beyond Machine Learning
Although Machine learning programs have the necessary characteristics to assist the trader in many tasks, their predictive potential has yet to be demonstrated.
Indeed, these programs only look at the past to try to predict the future. However, for this to be effective, the past must contain real lessons about the future and these lessons must be put to good use.
For this reason, research is now focusing more on the potential of Deep Learning programs applied to trading in order to study to what extent their anticipation capacities allow them to “beat the market”, at a time when these programs have already succeeded in beating the best Go players and e-sportsmen on the planet.
Hi Nicolas, I’ve tried supervised learning in Python (random forests) and sometimes results are ok but the problem I have is a) lack of large amounts of intraday data (for training) and b) even with a good result it’s not possible to see the logic/method/strategy that the machine thinks it has identified. I would be interested in using machine learning to create an adaptive strategy, taking a baseline from a “human devised” strategy, but then using the machine learning to consider how to adapt it to different market conditions, perhaps linear regression of optimisation variables. A good starting point would be if PRT walkforward analysis was able to output market conditions for each WF step. For example, average MACD line, ATR, ADX in every WF step. These outputs could then be used to look for patterns and initially manually fit to create adaptive algos, and then perhaps PRT could work out how to auto fit the data and pick best fit? I’d be keen to see this type of further development, but perhaps it’s already been done? Would be interested to hear your thoughts, many thanks
WF is a tool for YOU to analyze the benefit of optimization and on which period and with what division between IS and OOS periods. Therefore the platform cannot take for YOU the decision to auto fit with a particular setting.