Quant Analytics Algorithm Question: Is there value to an algorithmic approach that works out-of-sample regardless of market success?
In looking for the easiest to capture low hanging fruit of the High Frequency Trading world, I build an elegant algorithm that could be applicable to general purpose instruments. It works well on high volume NASDAQ stocks in the sense that it would make a profit in out-of-sample simulated trading. If you assume the ability to capture some reasonable percentage of rebates then it becomes very profitable (~200%/year). This seemed to be the type of results that I was looking for so I went live with this strategy for around 6 months. However, market dynamics (slippage, game theoretic behavior, market impact, etc) made the strategy unprofitable in live trading.
If I had another 2 years I might be able to learn enough market dynamic information to make the strategy profitable. However, after working on a poker startup and an algorithmic trading strategy for 3 years, I didn’t really have that luxury. I’ve been wondering if there is anything useful that I could do with this strategy. I believe that in the right hands with a firm that understands market dynamics, then this algorithmic approach could be useful.
So, here is the question. Does an algorithm that generates out-of-sample profits have any value regardless of success when it is crudely deployed to the market? Presumably, my out-of-sample profits are somewhat dependent on flaws in my simulation but to some extent that just means that my training runs also needed to better reflect market realities. The piece that I seemed to be most missing was game theoretic behavior of other market participants and a better understanding of how to deploy orders across a fragmented market.
I’d love to hear any opinions out there. Contact me privately if you see some value in this algorithm.
what platform did you use to do the back test? HFT algorithms are notoriously difficult to back test without a specialized back test engine.
One of my own creation. It seemed to use reasonable rules but …
fill assumptions in HFT are very tricky. I cannot count the number of HFT algorithms I have had work all the way through forward testing in simulation with what I thought were conservative fill estimates but when actually deployed live did not work.
The algorithms often missed by only a few milliseconds but there is no prize for 2nd place in that world.
You may want to consider something I have found to work; take the basic HFT algorithm and move up a time frame. If you were trading on milliseconds, consider seconds, etc. Many of the same techniques work if tweaked a bit.
FACEBOOK ACCOUNT and TWITTER. Don't worry as I don't post stupid cat videos or what I eat!