What is the CyBorg Predictor?

Monday, May 13, 2019

Tom Suter
Junior Data Scientist


This article is the first of a series aiming to provide our community with insight into the work done by the SwissBorg Investment team. We will be introducing machine learning and swarm intelligence applied to the investment space of cryptocurrencies with the final goal of creating a new enhanced bitcoin Index.

Predicting bitcoin trends is no easy task; you may wonder if it is not better to toss a coin. But instead of making you do that the Community app has, within the Daily Bitcoin Analysis available in the learn tab, two intrinsic indicators: the Community and the CyBorg Predictor.

  • The Community trend provides the dominant pattern of other players' forecasts over the last 24 hours: are the majority of forecasts "up" (bullish) or "down" (bearish)?
  • The CyBorg is a more complex algorithm that analyses the precedent prices to predict what is likely to happen next. Below, you can see how it can be built.

What is machine learning?

The CyBorg Predictor is a class of machine learning algorithm. It uses what is known as training data, the data for which we know what to expect, to learn. One common example is image recognition, where the model needs to predict the presence of a cat on an image.

In our case, the training data is made of timeframes of the historical bitcoin prices and the solutions: did it go “up” or “down”  24 hours later which would correspond to “A Cat” or “Not a Cat”. The model will train, meaning it will try to tune its own parameters to maximise the correct predictions on the given data.

Once we have this model that has learned, it is possible to give it a new data point. The model will use the found parameters to give the expected solution.

Similarly to us, the model performs better the more it sees data and the closer it gets to the new ones to predict. Imagine if cars start looking like planes we might be confused by labelling them cars.

So, if those programs learn from experience why aren’t we better? Well, computers can store a massive amount of information, and reuse it very precisely. Furthermore, they can come up with many combinations of data that we never fathomed.

For example, technical indicators in finance have been studied for years and investors have tested them in many market scenarios before qualifying them as good or bad. Machine learning can potentially use all the data to calculate tons of new “indicators”, test and update them to reach the one working the best in just a few seconds.

However, stock market data are hard to model. Unlike images, there is a lot of noise present in the data. Many factors influence the supply and demand, and therefore the trend which can influence the factors back. Capturing all of them is impossible. On the contrary, on an image, all the information is available to make the prediction.


How should I use the results from the CyBorg predictor?

As explained above, market data are incredibly complex, and a machine learning model might not be able to have a high certitude on the output every time. That is why it outputs a probability, and as it gets closer to 50%, fewer risks should be used.

In the Daily Bitcoin Analysis, you will find the predicted probability for the expected trend given at the end of the previous day for the next one. The probability should not be interpreted as the probability for the price to change in a direction but how the model interprets it. If the new data point is close to previous ones used for training, the algorithm will be more certain, thinking that the trend will be similar and the probability closer to 100%.

How does the CyBorg compare to me?

Each time a forecast is available, the CyBorg can predict. If unsure, it might skip and wait for more relevant information. The risks taken are matched to the ranges of the probabilities given by the algorithm. The current score of the CyBorg Predictor can be re-traced using the data from the beginning of the competition. We can, furthermore, calculate the failed and successful predictions and calculate the current accuracy of the bot which is the one displayed in the Daily Bitcoin Analysis.

We compared its score over time to the users’ scores. We considered only the users that had performed at least twenty forecasts, and among them, less than 1% are currently beating the bot.

Overall, the CyBorg predictor can enlighten you about the trends of the market. It is not an answer to what will happen next but it is a new indicator that performs better than most of your competitors. Combined with the other tools provided in the Daily Bitcoin Analysis, it becomes a significant boost in your climb up the app’s ladder.