We focus on powerful algorithms and modelling techniques that allow you to extract the most information from your data.
Machine learning, one of three subcategories of artificial intelligence, provides a basis for data mining. It seeks to the answer the question:
“How can we build algorithms that automatically improve with experience?”
Formally, it can be defined as an algorithm that is able to learn from an experience, with respect to some task, as measured by a performance metric.
In a trader’s terms, this an algorithm that looks at an indicator value (the experience) to make a trade (the task), with its success measured by the profitability of that trade (the performance metric). The algorithm then tries to learn what indicator values lead to the most profitable trades.
There are three general types of machine learning:
- Supervised learning:
With supervised learning, the algorithm is told the correct answer. For example, you would show the algorithm a dataset that shows each bar with the corresponding RSI value and whether the bar closed up or down. The algorithm then tries to learn which RSI values lead to price increases and which ones lead to price decreases.
With unsupervised learning, the algorithm does not know the right answer and tries to figure out the relationships in the data on its own. If you gave an algorithm a dataset with the values of an RSI, volatility indicator, and some measure of a trend, it could tell you that RSI values around 50 are related to low volatility and weak trends.
While not as popular as supervised or unsupervised, reinforcement learning is a combination of both. An algorithm isn’t told the correct answer but is rewarded if it makes the correct decision. It is usually applied to systems that must learn over time. If you had an algorithm designed to rebalance a portfolio, it would have to decide whether to add or reduce a position where each decision affects all future decisions.
TRAIDE employs a few different types of machine-learning algorithms along with both supervised and unsupervised learning techniques to find robust patterns in the data you select.
Data mining, a subfield of computer science, is the extraction of underlying, previously unknown, and valuable information from data.
There are 6 main techniques for data mining:
- Association Rule Learning
- Association rule learning is the method of finding interesting relationships, or rules, in a data set.
- The advantage of these rules is they can often be written in simple forms such as “if the RSI is below 30 and volume up more then 2% then buy”.
- Classification is the prediction of discrete “classes” of outputs.
- For example, a classification algorithm would predict whether the market will move “Up” or “Down”, as opposed to the magnitude of the move.
- Regression, in contrast to classification, is the prediction of a continuous, numeric value.
- It would output the direction and magnitude of a market move, for example “+.017%.
- Clustering is the process of dividing data into similar groups.
- It can be used to asses risk by identifying which investments are more similar than they initial appear.
- Anomaly Detection
- Also known as, “outlier detection”, anomaly detection identifies data points that are abnormal to a data set.
- In the financial industry this is usually associated with fraud detection.
- Summarization involves techniques for finding a compact description of the data.
- This can be everything from simple metrics (mean, standard deviation, etc.) to complex visualizations and diagrams.
TRAIDE uses a unique combination of association rule learning with robust classification and regression techniques. We use our own proprietary algorithms designed to find the underlying patterns in your data and the “Strategy Dashboard” allows you to translate those patterns into easily understood rules.
Association Rule Learning
Association rule learning is the method of finding interesting relationships, or rules, in a dataset. These rules are usually written as “if/then” statements such as: “if volume is up by .02%, then there is a 65% chance the asset price will rise in the next 4 hours.”
Most association rule learning applications rely on an algorithm, like the Apriori algorithm, to find the rules for you. However, this process has several drawbacks when applied to building trading strategies.
- You must clearly define a single metric to define the relationship.
This could be net profit, return per trade, drawdown, but not all three. As a trader, you know the importance of analyzing more than one metric of your strategy.
- An increase in complexity.
- There are thousands of different technical indicators you could have the algorithm explore, and that’s not including fundamental or sentimental data.
- Once the algorithm has analyzed the data, it is very difficult to know where these rules are coming from without having an in-depth understanding of exactly how the algorithm works.
- It becomes a “blackbox”.
- There are also many different parameters that must be optimized for algorithms like the Apriori in order to run a robust test. You must have experience working these types of algorithms or you greatly increase your chances of overfitting and finding spurious correlations. These correlations are much less likely to hold up moving forward then rules with a sound theoretical background.
With TRAIDE, your using your experiences along with the patterns found by a powerful algorithm. You select the combination of metrics that are important to you, the indicators and assets you want to explore and we ensure you are running a robust test.
We’ll show you the results and you finalize your rules into your own trading strategy.