I agree with Cobgoblincave4 that we are hitting a bit of a bump in the road with the accuracy figures for XRP; currently sitting at 75.54% when forecasting price volatility. I am confident though that we will get it up into the 90's like many of the other tokens we cover.
Latest forecast and historical for XRP below as an example.
So how do you improve the accuracy? Do you feed new data or is there any tweak to the algorithm? I believe some coins are just wildly volatile since their fundamental changes quickly like XRP after the SEC case, so it is not surprising if your current model failed to stay within 90% like other coins. People should be aware that this 'prediction' can't correctly express any change in fundamental values if I get your explanation correctly.
Good question. First off it should be noted that such a undesirable dip in accuracy will not cause a kneejerk rection from the team resulting in unnecessary fiddling with the model. So for example something like an undesirable outcome of an SEC court case can tank the price of a token out of nowhere and also catch the ML unawares thus driving its accuracy percentage down. We would simply leave the model to readjust its forecast to the "new normal".
GNY maintains a data warehouse containing all available historical price information for the twelve supported tokens and twenty five of the most popular and predictive chart analysis trading indicators. Our Fabian model is trained and tuned using this historical data.
On the first of each month, Fabian automatically conducts a hyperparameter sweep in order to tune itself and ensure that the model is performing to its maximum capability and efficiency.
Compared to the previous machine learning model, Fabian utilises more data points and those data points that it uses have greater granularity. For example, the original model used 1 day candles and indicators whereas Fabian uses 6 hour candles and indicators. This allows Fabian to be more responsive to predicting the likelihood of short term volatility and relative value changes. We are continually adding to this curated pool of data points and we believe this can lead to more accuracy and further-looking forecasts.