This is Part 6 in multi part series:
- Part 1: Basic strategies, introduction, setup and testing vs June-July market.
- Part 2: Advanced strategies and where to find them, testing vs June-July market.
- Part 3: Basic and Advanced strategies testing vs August market.
- Part 4: Neural Network strategies description and backtests against September market.
- Part 5: Neural Network strategies backtests against October market.
- Part 6: Did Neural Network strategies predict November 14th price drop?
- Part 7: Crypto Trading 2018 in Review: 17 Advanced + 15 Neural Net strategies tested
- Part 8: Intro to Statistical Arbitrage in Crypto — Pairs Trading
- [NEW] Part 9: Crypto Trading 2019 Half Year Review: 17 Advanced + 15 Neural Net strategies tested
In this part, I wanted to get started on whole “Neural Net prediction accuracy” thing I promised in last part, by backtesting a big recent event — dump that started on Nov 14 and has been going on about 2 weeks now. If we count from Nov 14th, here’s how the situation looks like:
- WORST % — worst % loss counting from Nov 14th 00:00
- CURRENT % —loss counting from Nov 14th 00:00 until Nov 26th 00:00
As you can see, this is insanely bad. One of the worst drops ever for Crypto, at least psychologically, because the overall mood in community was that November will be BULL month, just like last year.
So in this part, I want to backtest the drop starting from Nov 14 (+giving NN history to learn) with different candle sizes and see if any of 15 Neural Net strategies predicted something useful.
If you are interested in strategies themselves and missed the last parts, here I have a quick description of each strategy, what NN libraries it uses and how it makes long/short decisions.
Note about ManuNet strategy: it’s the only one, that doesn’t predict price — instead it predicts price change direction (up/down) with threshold. To display the predictions, I will draw 2 scatter charts at the bottom — red dots will represent
short predictions, but the green —
5m candles and 900 candle warmup
Let’s start with 5m Candles + 900 candles as a warmup period (for Neural Net training). That gives us 3 days as history. Could be not enough to learn any patterns, but will see.
Here are the results. Green line is the prediction made by strategy.
Few things that we can notice here:
- neataptic / mounirs_esto / neuralnet_zschro — closest fit to the line, but not in a predictive way. Just reacts quickly to what has already happened, that’s all.
- LSTM_MACD_RSI_V3 definitely the craziest prediction line from all strats, I think it’s related to
this.divisor = 10000, I think it should be more dynamic value, instead of hardcoded.
- ManuNet — looks interesting, screams SHORT all the time (which is correct), but the few LONG calls seem to be quite reasonable.
- mounirs-ga-version-1 — there seems to be some bug in strategy, returns straight line once the
check()starts running (warmup period ends).
10m candles and 5000 candle warmup
Ok so maybe 900 candle warmup wasn’t enough. Let’s increase candle size to 10m and bump the warmup period up to 5000, which gives us 1 month+ of training data.
I don’t see much difference, except that LSTM_MACD_RSI_V3 got even more wider range of predictions.
60m candles and 2000 candle warmup
Ok last try — 60m candles, but let’s reduce warmup to 2000 candles, which means almost 3 months of patterns to learn. Let’s not go deeper into the past, because that could mean NN learns old patters that are no longer actual.
The picture looks different but still — I don’t see any predictions going on … except for ManuNet in some interesting cases. Take a look at the last drop at Nov 24/25th, predictions go red few candles before the actual long red bar. There’s a single green one between, which would still mess up your trading in this case (because you go short/long/short and pay commission), but at least this looks like a prediction. Also, the diagonal drop from Nov 18/19th shows red dots early, but this one is arguable. Other than that, I don’t see much predictions going on.
Doesn’t look like Neural Nets were of any help in this situation. I didn’t hope they would predict the start, but once the 1st wave was over (Nov 14th) I hoped that strategies would at least catch Nov 19, Nov 20 and Nov 24, but looks like this is not the case. Of course, predicting such rare events are not easy and might not even be possible.
In next part I’ll dive more deeply into why we see such predictions and how can they be improved.