In the course of solving the problem for predicting the BTC / USD bitcoin rate from historical data, an understanding came of what should be a computing system with which you can build the desired model. I wanted to describe this vision for the respected Habr-world and in this way to understand how really interesting it is for the current reality.
So, we consider some problem in the statistical setting. This means that we have some data set, for simplicity, the price of BTC / USD from time to time. Suppose we set the following goal: to build an estimate on the average value of the BTC / USD price the next day, such that the specified quality functional is minimal. In the simplest case, as the quality functional, we can take the sum of the deviations of the actual average value from the predicted value, taken modulo and normalized to the actual value.
As it turned out, and not surprisingly, it is relatively easy to build a model for prediction. Even more precisely, as a rule, a huge mass of models appears for such a problem, which can differ both in the value of the tuning parameters, and in the number of internal degrees of freedom and other internal structures. Therefore, the solution of the problem is reduced to the following problem: it is statistically correct to choose the best model from the available set of models.
To solve this problem, we need a system for statistical inference, which you want to have:
- you need to specify a model in the system, describe the tuning parameters and specify the data on which the model will be evaluated (trained). The data used will not be used in testing this model;
- for many models, you need to specify a quality criterion through the calculations of which the models will be statistically compared;
- when comparing models, it is necessary to work out the criteria for false-positive / false-negative conclusion (when the data for the conclusions are either not enough or the comparison is impossible due to statistical error).
The main feature of the system: its conclusions will be statistically substantiated (within the available information) and, in accordance with mathematical statistics, will asymptotically converge to a truly true value with an increase in the number of independent data.
Offhand, this system will be useful for a large class of tasks, including for testing models built using Machine Learning technologies.
And if...
- if such a system is already available and accessible (with realistic learning curve ), then you need to run to use it;
- there is no such system yet, then you need to sit down to do it.
Afterword
- As often happens in life, the true size of the complexity of the task becomes clear when you plunge deeply into the problem ...
- An example of a model for predicting the rate of Bitcoin BTC / USD the next day can be found here .
PS:
You-dy-sch: yeah, you-dy-sch ...