Risks in Machine-Learning Searches for Intraday Futures Strategies

# Is my trading strategy search methodology sound?


# Is my trading strategy search methodology sound?












I'm building an algorithmic trading business. I'd be grateful for informed comments and opinions on my trading strategy search methodology.

Goal

Develop (profitable!) fully automated intra-day trading strategies.

Markets

I'm focusing on FX futures and equity index futures on CME and LIFFE. Becuase:

- I know these markets best

- I can get roundtrip times of 10s ms

- Futures margin means big leverage for my small account

Data

I'm buying in Level 1 tickdata from the exchanges. I don't think level 2 data would add anything because I wont have low enough latency to make use of it. A typical contract might have 1M data points per day.

Search Strategy

I have a bias towards machine learning specifically SVM/R. My plan goes like this

- Choose an instrument and forecast horizon (5-120 seconds)

- Assembly a large stable of features e.g. Resampled lagged time series and returns Wavelet transform of the above Measures of information entropy Some indicators from Operators on Inhomogeneous Time Series (This is 11 years old now but I've yet to see anything that meaningfully expands on it) If can hold down the vomit some 'Traditional Technical Analysis' indicators

- Choose a subset of features with genetic search. The objective function is the MSE of the n-fold SVM cross validation

- Build the final SVM with the best parameter selection.

- Hope that the highest confidence predictions form the basis of a profitable trading strategy

In your experience do you think a strategy developed like this can hope to make money on liquid equity index and FX futures? If not why not?

## Answer by Tal Fishman (score 21)

https://quant.stackexchange.com/a/2149

This is a very interesting question. I believe it is getting a lot of up-votes from people who have wondered the same thing and don't know where to begin, whereas you have at least laid out a reasonable-sounding plan. I commend you for that. However, it is not clear to me what you're trying to learn by posting this question. In my opinion, the plan you laid out seems like a reasonable start but, by itself, will not lead you to a profitable trading strategy. That is just my opinion, is there something else you'd like to know? The rest of my answer consists of further comments.

If you haven't already, I recommend you read Ernest Chan's Quantitative Trading. He devotes an entire chapter to "Fishing for Ideas: Where can we find good strategies?" I should mention, however, that he is skeptical of your bias towards machine learning:

> At the risk of oversimplification, we can characterize artificial intelligence (AI) as trying to fit past data points into a function with many, many parameters. This is the case for some of the favorite tools of AI: neural networks, decision trees, and genetic algorithms. With many parameters, we can for sure capture small patterns that no human can see. But do these patterns persist? Or are they random noises that will never replay again? Experts in AI assure us that they have many safeguards against fitting the function to transient noise. And indeed, such tools have been very effective in consumer marketing and credit card fraud detection. Apparently, the patterns of consumers and thefts are quite consistent over time, allowing such AI algorithms to work even with a large number of parameters. However, from my experience, these safeguards work far less well in financial markets prediction, and overfitting to the noise in historical data remains a rampant problem. As a matter of fact, I have built financial predictive models based on many of these AI algorithms in the past. Every time a carefully constructed model that seems to work marvels in backtest came up, they inevitably performed miserably going forward. The main reason for this seems to be that the amount of statistically independent financial data is far more limited compared to the billions of independent consumer and credit transactions available. (You may think that there is a lot of tickby- tick financial data to mine, but such data is serially correlated and far from independent.)

The primary reason I believe your plan is incomplete is that you haven't demonstrated an edge. All the techniques you mention, machine learning, inhomogeneous time series, wavelets, are not new. If you can demonstrate a novel yet useful way of combining these methods, then maybe you have something. But be aware that there is huge competition in FX and index equity futures, and it is extremely unlikely that no one has tried something along the lines of what you are trying.

Also, just because you think you know a market best does not make it the best market for you. For that, I point you once again towards Chan's book. He points out that your strategy should match your

- working hours

- programming skills

- trading capital

- personal goals

Think carefully about each of these. Note that prior knowledge of a given market is not on the list. That can be picked up relatively easily. In fact, given that you already mention you have a small account, I really see no reason for you to be in such liquid markets, particularly as getting the sort of ultra-low latency to beat other market participants in a market-making type strategy will take an investment of many $100k.

## Answer by Hasselhoff (score 0)

https://quant.stackexchange.com/a/80331

I'll make the recommendation that you are definitely going to need some domain expertise for an SVM on financial data. Whatever your experience is with machine learning outside finance, expect it to break in financial data. Data generated by markets is quite specific to markets and not like natural processes at all. Not to say machine learning isn't effective, just needs a domain specific approach.

Brush up on your time series management particularly forward looking bias in your back testing. There are ways to manage the dependency in financial data (mentioned in the answer above), but you can't survive forward bias. If you are going to do this, you'll need to rise above "data science" and be a statistician and really understand your data. Your strategy cannot be a black box. You could even say that algo trading is not the trading business, it's the data management business.

Always consider trading friction in your back test. Friction can turn an all the money in the world strategy into a total loss. A lot of winning strategies are little more than better pricing, or just plain price arbitrage.

Those are two major pitfalls, but there's a lot of them. Good luck!