System Development

This thread is dedicated to the development of a few ML(Machine Learning) systems I am working on. The main purpose is for me to record the results of the systems as I am currently developing them. Once I am satisfied with how they are performing in the backtest I will move to forward testing them on a live account and that will also be tracked here. I will not be diving into the logic behind the systems as this is not what this thread is about. It is strictly to journal my progress in developing the systems.

Systems:
1. Low Strike
2. High Strike

Brief Description of the Low Strike System
This system is fairly simple. It trains itself on the distribution of price over the last x amount of bars. It then makes a prediction of a suitable entry price, stop level and profit target for trading the next bar. If the profit target or stop level is not hit during the next bar the trade closes at market price when the bar closes. It retains itself on every bar to adjust to the market as it changes. It is called the Low Strike system because it is not meant to predict the market with extreme accuracy. It is meant to predict the market with enough accuracy to make a profit based on the R:R it defines. I am shooting for the algorithm to predict trades with 30-40% accuracy, hence the name low strike. The R:R ratios generated by this system range anywhere from 1:3 to 1:10+.

Brief Description of the High Strike System
This system is also quite simple. It functions very similarly to the last one except that it models the price distribution and makes predictions over huge amounts of bars. It also retrains with every new bar and adjusts the orders accordingly. Because of this, it is only traded on the 1-minute chart. Despite this, the trades can last anywhere from a few hours to 20 days+. The goal of this system is to predict the market with near pinpoint accuracy(95%+ Win Rate). Because of this the R:R ratios generated from this system can be anywhere from 1:0.10 to 1:10.

Very interesting Sis. I do like a bit of machine learning! Of course your systems must assume that past price behaviour is predictive of future price behaviour, which to some people is controversial, but personally I think that this is true more often that not. Why? Because humans drive the price, either directly or via computer programs they've written. And if there's one thing that remains constant over time, it's the fear, hope and greed of humans.

Anyway, the proof is in the pudding so I'm looking forward to seeing your results.

By the way, there's a typo in your second post. The second title should say "High Strike" not "Low Strike"

System: Low Strike
Pair: EURUSD
Spread: 3 Pips
Testing Period: 01/01/2010 - 06/30/2016
Bar Size: 240
# of Bars Trained On: 200

Thoughts: Wow... Total burnout on this timeframe with the above settings. The biggest contributor is the spread. The trades are not that big Pip wise so a 3 pip spread is extremely hard to overcome, considering some of these stops are and targets are within the 3-20 pip range. Although this spread is really high for the EURUSD it is always interesting to see how the system will perform in terrible trading conditions.

System: Low Strike
Pair: EURUSD
Spread: 0.2 Pips
Testing Period: 01/01/2010 - 06/30/2016
Bar Size: 240
# of Bars Trained On: 200

Thoughts: Still nothing special. Although it did end up positive, equity was very unstable the whole time and the results appeared to be quite random. I still think a big factor to these lackluster results is the spread. Even though this test it is very small, it is still quite large %wise compared to the amount of pips we are trying to snag on each trade. Trades are based on volatility so when the stop and take profit levels are tight the transaction cost are anywhere from 5-10%+... Much too high. Something else to keep in mind is even though the program is trading a 4 Hour chart, it is trading the chart one bar at a time so the trades are actually quite short term. Another thing I suspect is holding this algorithm down with the above settings is the number of bars it trains on. I think it is overfitting to the market which makes it really hard to adjust when market conditions change very quickly.

System: Low Strike
Pair: EURUSD
Spread: 0.2 Pips
Testing Period: 01/01/2010 - 06/30/2016
Bar Size: 240
# of Bars Trained On: 25

Thoughts: Using fewer bars to train on, we can see the algorithm performs much better as it can react quicker to the changing market conditions. Although the performance is still pretty lackluster with a majority of the time the account being in a drawdown period, the performance has drastically improved as the balance remained positive nearly the whole time and the run-up was over 200% higher than when we train on 200 bars and the account ended with a higher balance as well. This leads me to believe that overfitting with this algorithm will be a big problem and it will be important to keep the training bar size relatively small so the system can adjust quicker. I still need to figure out what is going on in the large periods of drawdown but that will be done later. First, I want to run this system over basic settings and timeframes to create good benchmarks and then will build upon that. Although these results are much better than our previous test, I still would not trade this system as it would be very hard to guarantee a 0.2 pip spread on every trade and eventually the transaction would likely catch up to us.
NOTE: I'm not going to share the results of this system with a really high spread because we have already determined anything above a super competitive spread will basically kill this system on this time frame.

I agree that to some extent, the past price can predict the future price for all the reasons you listed. Even if the market is truly random, there will always be some sort of pattern in the distribution of the data that would be tradable for a short amount of time. The trick is that nothing will work all the type so you need to constantly be adjusting to the new distribution of price. Also thank you for pointing out the typo. It is now fixed! Please let me know if you find anymore.

Here is a great video that touches on this distribution of randomness a little bit for anyone who is interested.

Thanks for the video, it looks great. I'll watch it tonight hopefully.

By the way, one quick question - is the x-axis of those graphs the number of trades? If it is number of trades then it looks like you're taking a trade on almost every bar. Trading costs (spread/commission and slippage) usually kill retail high-frequency automated systems, so it might be worth extending the length of time you're in a trade to more than a single bar. In other words, wider stops and targets to reduce the cost per trade.

That is correct, the system trades every 4-hour bar(with the settings tested above) as it is predicting how price will move for the next 4 hours and if the stop or take profit level is not hit during that candle or bar, it closes the trade. You are right, the cost of trading is definitely what is really killing the system trying to trade it on this timeframe. It is not necessarily because the amount of time the trade is opened small, it is because of the volatility that is to be expected in that amount of time is small. Because the volatility of each 4 hour candle is relatively small, the cost of trading becomes extremely expensive and as seen in the results above, so expensive that it can not be overcome consistently.