I created a new algorithm from scratch that predicts the bitcoin price of bitstampUSD. Preliminary testing yields results around 200% gains per year on top of market fluctuations and initial investment, not accounting for trading fees, with one trade per day. I'm looking for someone willing to invest or do stock sharing with this algorithm. Here are some preliminary figures. First: Training Error by Iteration (during the model creation process) Second: Predictions (Green) vs. Actual Deltas by day Third: Trading advantage by day (Trading advantage = advantage over squatting on a Bitcoin investment) (X-axis = days since January 1st, 2016, Y-Axis = USD) Note: These predictions are created from and tested on the same dataset. Some rounds of simulation have yielded $800+ gains on top of a $430 initial investment, and the performance should improve with time with more data points due to the model features involved.
If your predictions were created and tested on the same dataset how do you not have 1000% returns? You absolutely have to have out of sample testing or you're proving nothing more than your ability to fit data. If you haven't taken a couple advanced stats classes I highly recommend you do since it looks like you're at school.
Since the models only use a small number of features, it's impossible to crank out arbitrarily high returns. But you're totally right, I should have done cross-validation before posting this. The models are selected based on trading performance, but they're NOT trained on trading performance, only on prediction accuracy. That is why the returns are so significant - the algorithm doesnt just run tons of random trading choices. These models are fit only on a very small number of parameters, so the results are still remarkable. Additionally, out of all the tested features, a couple of them show up in all of the feature subset variations. Another pressing point, the model is only fit on about 365 observations, unlike other models which need more data and processing power. This is because the particular model is extremely robust and accurate compared to standard multiple regression. A final point: I haven't fine-tune the model yet to collect accurate data, thus why I haven't performed cross-validation. These are only preliminary results and again they're fit on only a small handful of weights at a time. And just to drive much point home, look at the BITSTAMPUSD 1-year market history chart. All the points of rapid gains perfectly coincide with the crashes and booms. That means I've got an intelligent algorithm here.
It seems like I've got some interest in the algorithm. I don't have much free time on my hands but I'll post the cross-validation sets since I know it'll be somewhat worthwhile.
Your training and test set are the same?! Sorry but that means your model is actually worthless. Having a separate training and test set is like the first thing a machine learning class will teach you.
Seems like a standard blueprint here in ET: 1. A poster shows some "unimaginably" high profits, and asks for money. 2. When asked a question, the poster says that he "doesn't have much time".