Thank you everyone for your interest. I have some extra time this evening so let's move our Github repo forward a bit. How shall we begin? When I last left off, I had already coded a downloader for OHLCV data. It uses the CCXT library to fetch market data from any arbitrary cryptocurrency exchange/symbol/timeframe. That script is here: https://github.com/TheDoctorAI/AI-Trader/blob/master/download_market.ipynb
Running These Scripts
Machine learning has seen amazing innovations recently. One such innovation from Google is called Colaboratory. When I first started learning AI, I bought a super expensive GPU so that I could train my models quickly. Then, Google went and released a free service that allows anyone/everyone to use Google's GPUs & TPUs (Tensor Processing Unit) for free! That's just... Wow. What a lucky time for people getting into AI. It also means that I can easily share these models with you all and you can run them on expensive hardware in a few clicks at zero cost. If you visit any of the .ipynb files in the Github I linked here, you'll see an "Open in Colab" button at the top:
Note: If you're going to run my AI models, don't forget to (In Colab, for each ipynb example) click
Runtime -> Change runtime type and then select
GPU under
hardware accelerator. No need for TPU at this stage (It won't make it faster). GPU will. Also, I have coded these so that they save and load data from Google drive. You will probably have to authorize the script to load the .csv file. Google will step you through that automatically when you run the notebooks (It's easy).
Our First Model: Multidimensional LSTM Bitcoin Forecast
We have a long way to go but we will start where I imagine everyone invariably does - LSTM (Long/Short-Term Memory) neural networks. The power of LSTMs can't be overstated. An LSTM is a recurrent neural network that incorporates a memory mechanism. If a neural network were a computer, the memory module of an LSTM would be the RAM & memory controller. I'm not going to get into the particulars of gradient descent here but suffice to say, LSTM solves some important problems that makes it useful for this task of forecasting a market.
Most Bitcoin LSTM model examples you will find only forecast the close price. Well, that's pretty lame. Any serious effort should attempt to forecast the OHLCV (Open/High/Low/Close/Volume). There are also technical indicators to consider (Moving averages, Ichimoku, Bollinger bands, etc). By training on multiple vectors/dimensions, we can incorporate a relatively arbitrary number of inputs. We won't be adding any technical indicators just yet but just know that we plan to investigate that. I especially have some ideas about using technical indicators and convolution down the line.
I have directed our downloader to grab 1h OHLCV data from the Bitmex exchange and save it to CSV. We then load the CSV into
our model and out comes our forecast:
Very good. We can see that we are fitting to the data nicely. The LSTM has learned the general patterns of Bitcoin and has now forecasted. I had to limit the amount of candles in the dataset to less than 400 due to an annoyance in the model that cropped up when adapting it to BTC/USD. We'll deal with that later. It is important though as we need to see how much data we can benefit training on.
What's exciting here is we can see that not only is our model projecting our fitted line into the future, it's also learning Bitcoin's movement patterns. This forecast has Bitcoin running back up to about $5,550 and completing the classic M pattern we so often see on the chart. This is an impressive intuition on behalf of the neural network. Of course we can argue about the complexity of the model but still, the resulting pattern is undeniable. If I were charting on my own right now with my own intuition, this is the sort of pattern I would project. Again, very impressive. We still need to do some accuracy tests. More to come on that.
Note: Anyone wanting to continuously use this forecaster, all you have to do is run the downloader notebook and then run the forecaster (Both on Colab). The model takes about 5 minutes to train. I have also added some code that prints out the forecast vectors at the end of the script (So you can do what you want with it).
What's Next
Alright, so. We need some more code for evaluating accuracy. We also need some more models. mLSTM is far from the best we can do. Very far. It is however a good start. Also, this is not a trading agent/bot model. We can't backtest it because it doesn't optimize for actual trading. For that, we'll need a different algorithm. Of course models like the one we just used can contribute to an agent's understanding.
Links
The market loader: https://github.com/TheDoctorAI/AI-Trader/blob/master/download_market.ipynb
The resulting csv: https://github.com/TheDoctorAI/AI-Trader/blob/master/data/btcusd-1d.csv
The AI: https://github.com/TheDoctorAI/AI-Trader/blob/master/lstm_forecaster.ipynb
Much more to come. Stay tuned.
