citlacom
commented
3 years ago Hi @thiloyes - FYI I published the machine learning models predictions in this repository at https://github.com/financial-astrology-research/financial-astrology-stats/tree/main/machine_learning/predictions I just keep the ones that have performed stable with a minimum average 3 months accuracy of 50% so we can track the performance for few more months before the final cleanup. Also I published the performance reports of all the models I have trained at https://github.com/financial-astrology-research/financial-astrology-stats/tree/main/machine_learning/performance There are multiples reports based on the day that was generated so we could analyse the accuracy decay after some period of time.
The most recent report is from today 2020-01-24: https://github.com/financial-astrology-research/financial-astrology-stats/blob/main/machine_learning/performance/models-predict-performance-2021-01-24.csv The rows are sorted by rank and I have removed the weighted average to don't weight more the current month accuracy. The performance report script is also published at: https://github.com/financial-astrology-research/financial-astrology-stats/blob/main/modelsPredictionsPerformanceReport.R
I noted that there was a bug in the categorical price effect that was using the difference between two moving averages instead of the daily price change so all the frequency stats tables are invalidate and needs recalculation, see: https://github.com/financial-astrology-research/financial-astrology-stats/commit/7eacb823dbd4caa3f1b3a9d9ce97e5a2172c84a6
I will process the recalculation and publish the new frequency statistics using the fixed categorical price "buy/sell".
During financial astrology mundane aspects research I have explored different machine learning models using R caret package http://topepo.github.io/caret/index.html and encoding the mundane aspects in different forms so machine learning models can generalise from past date what may be the expected outcome of aspects that are not yet seen for a relatively new trading asset like cryptocurrencies which only have been traded for 9 years or less. Many of the slow planets aspects like the Saturn square Uranus that is currently forming have not occurred within the price historical data so exploring different ways of aspects compression can extrapolate the effects by aspect type or slow planets aspects activation.
Multiples models have been tried using this encoding tricks of the daily mundane aspects and trained with different algorithms such as: KNN, Ridge, GLM, Lars, Lasso, Enet, SVM, RRF, Rborist, Pls, Mlp, GlmNet, GAM, Gauss, BayesGLM, and MLP with different hidden layers architecture.
The best fitness was observed using KNN with K set within 6 to 10 range and using cross validation and ensamble of multiples trained instances of the same model to compensate the random nature for the CV training that may return different results over each training attempt due the data set split randomness.
Some models have achieved about 70% accuracy in predicting the daily price direction (path of leat resistance) during a period of 60 or more days of unseen data so our group member Thilo a professional mathematician see some possibility that this results is not just random. A more complex simulation will be required in order to do some scientifically conclusions but this is an interesting finding that give us some filter within the financial astrology rules search space.
The models predictions that was mostly generated in November 2020 will be included in this repository under "machine_learning/predictions" so more people can contribute with probability analysis to conclude if any of this models could achieve this level of accuracy just by chance.