How’s the profitability of your first approach going? Still running with it live?
Hi, sorry for the late reply.
So, all in all it is running on auto and going well. It makes more than I did with my full time job.
Here are some interesting things to come out of it:
1. It doesn’t use any of the indicators that are taught for trading. They are pretty much worthless.
A while ago I added something interesting to the backtesting system - random trading.
imagine a strategy gets 200 signals a year. The system would make 200 random trades as well, randomly selecting buy / sell with a trailing stop.
To my surprise random trading was not that bad - often only a few pips lost in the year, and more often than not out-performing whatever was being tested.
This was where I started to realize how bad the ‘indicators’ are.
Here is an example: I changed the random trading to using the EMA 100 / 50 / 20 (or whatever) to identify and trade with the trend. So it bought when above the EMAs and sold when below (with EMAs in order).
This performed worse than simply random trading.
I applied the same to all sorts of indicators - it’s quite eye opening, and quite hard to beat random trading.
The bot works on price action alone.
2. Overfitting is an issue
Often you read ‘X trategy only works on Y pairs’. This is bullshit. You’ve simply overfit your strategy.
The system tests all the pairs / CFDs and indexes it can find. It takes into account spread, and chops off trades with high spreads, though.
Overfitting is very easy to do.
3.Some crazy days have happened
A few days ago, when Trump was announced with Covid, the bot was 18k€ up in about three hours. This dropped off over the next couple of days, but it was still a good week.
Out of the last eight weeks, I’ve had one losing week and the others are all profitable.
4. Reliability is the biggest issue
My biggest issue has been REST API calls and errors - so data not coming back or connections blocking. The system has about 10% code for trading and 90% for logging an error handling.
5. So called fundamental analysis is still more profitable, for me
I still think it’s better just to stay vigilant.
Two recent examples:
- Boris Johnson announcing stalling of brexit talks. That was a Monday morning a couple of weeks ago. I shorted the GBP all over the place and took a lot of profit.
- Trump got Covid - again. shorted the markets and took a lot of profit.
At the moment, the markets are moving huge amounts accoring to news from US, EU, UK and Japan.
HTH!
2 Likes
Glad to hear that it sounds like your development efforts have been successful lately. There may be hope after all.
Recently, I have worked on running some Monte carlo simulation against my money management algorithms with ‘random’ entries programmed at 50%. It seems that my current money management algorithm would help reduce drawdown and make profits more explosive during streaks of losses and wins, respectively, but in the end it still achieves the same return level… and ultimately profitability still depends on the ‘edge’ of the system. That is, at the 50% level, it would still be a crap shoot whether you are up or down in the long run.
Do I interpret that the ‘grid’ system that you use for entries & exits, that is, price action alone (if I have mis-interpreted, let me know) , and using NO indicators in the system is what has given you the recent profits? (or another type of price-only system, whatever that is)
I have recently started development on a machine learning project in which I hoped to feed multiple technical indicators and gather a simple prediction on whether price will achieve some distance. Doing this type of research I would hope to increase my ‘edge’ , so that I could still use the same money management algorithm, but enhance the longer term results. But I wonder if you think this wouldn’t even be a worthwhile project in the first place given your take on technical indicators? Any take on machine learning based technical analysis?
I agree that ‘raw’ indicators likely do not work as they are shown in many of the beginner materials… it seems some say only use them under the right market conditions… still seemingly wizardry like analysis techniques to me…
Also I have observed the same problems using some broker REST API . Frustrating.
Cheers
Hi,
The grid system never worked in the long run, either live testing or simulating. 
Regarding the ML stuff… be careful of one thing:
The most important part is selecting the inputs for whatever technique you are using - they must be independent of each other.
Let’s say you choose RSI(14) and ATR(14).
Both are derived from the last 14 candles - in other words you are using the same information twice.
Better would be to use the pure candles. Same for MAs, ADX etc etc - they are all derviatives of the same raw information. This leads to crap results and overfitting.
However, it raises another issue: The candle prices are also related to each other, and are not independent.
So we have a problem - and this is often why you see people downloading tensor flow, throwing indicators and prices into a Neural Net or SVM and saying “doesn’t work”. Or even better sayying 'Only works on GBP_JPY on a Tuesday morning".
You need to find a way to manipulate the data such that it conforms to the requirements for the models to work.
Then you can start looking at ML.
And you’ll see some success, in my experience.
Thanks for the insights!
You’ve reiterated the importance in my mind of the “feature engineering” part of ML projects.
I understand your point about independent input information. Although, I always thought that providing the algorithm with certain specific indicators would basically reinforce to the algorithm that ‘this type of derived infomation is important’ as in the case of maybe an ATR14, because you’ve explicitly told it pay attention to this volatility feature.
More experimentation to be had on my part. Thanks again. I can toss out the ‘grid’ system now
Yes, but if you think about it - the ATR is already in the candles themselves.
So let’s say ATR was the magic indicator - a SVM, for example, would classify the trade decision just as well with ATR as input or the prices themselves.
On its’ own the ATR would be ok, but as soon as you start mixing other inputs as well you’re effectively sending the same price information twice.
And still to note: I don’t think that’s the right way, anyway, as each price is related to the other in some (holy grail) way. But ymmv
Another issue with ML is the number of trades that are good - let’s say in a set of 1000 candles you have 100 good trades.
You find that your model trains on the training set, and then gives 90% accuracy on the test set.
And then you realize it just said never trade 
I think there are better ways to attack this, there was a post on reddit recently that provides a really good way into some good strategies.
I’m not sure about links, but it was here.
Brief guide on researching strategies and generating alpha
This was similar to the journey I took
Re. ML I’m bumping into similar case where it seems the neural nets I am trying are finding local minima too quickly and always making prediction of the class with the most frequent cases… very similar to your example.
Other research I’ve been wanting to try on price data are the Digital signal processing techniques like extracting the cyclic nature, waves, phases, etc. Perhaps use those as input features, but I’m not convinced that additional derivation of the same price data would be of any use to the model, as you’ve pointed out, because it would contain the same first-order information
That’s a really interesting reddit thread. Thanks for sharing that 
An aha moment perhaps in focusing on pairs trading/stat arb/mean reversion and not simply the price data of only one instrument at a time. Seems to lend well also to your price only data, but used in different ways.
There are a lot of traders who prefer automated trading. But as a trader myself, I like to trade manually and become the master of my own trades.
I know about automated trading but have never been able to try my hands at it. Looks like now is the right time.
My automated trading experience has been great till now.
If you want to try your hands in algotrading, there is a software called ApiBridge that allows free papertrading for you to try and use their features. You can try that. Their website link is: Home Particle Animate - Algoji . I personally used it and liked the interface as well as its execution speed. Go and try your hands in algo trading yourself and see if it suits you.
Do message me if you need any help.