Is there something wrong with optimizing trading strategies. I think the whole conception of optimizing may be a wrong paradigm. In fact during optimization we make in the same time the strategies very vulnerable to randomness.
That is the whole point. All trading strategies are very vulnerable to randomness but the problem is that y we do not know to what extent they are vulnerable.
That is why the use of any fancy digital filters or enigmatic time series preprocessing methods: SSA, wavelets, PCA does not add much.
On the other hand the use of different optimization algorithms does not add much too: back propagation, genetic algorithms, particle swarm optimization etc.
So what is the reaction of the big guys. They were investing in infrastructure in order to use the infrastructure edge together with complex data mining methods. And that worked and works. However as many others will do the same the competition will increase in the low latency market. Even more some countries are planning to tax the HFT, in France this is already a reality.Big boys are also using traditional macro strategies but this is another beer.
My point here is that a successful trading strategy needs to balance two pillars:
a. An adequate optimization regarding the market conditions. In practice that means that the strategy needs to learn from the market conditions and not to over - fit the past market conditions. This is a hard stuff.
b. The randomness robust trading strategy. There is a need to secure that the trading strategy is robust regarding the randomness of the market.
There are different ways to achieve it.
- To measure the state of predictability of the market and trade only when the market is sufficiently predictable. This is the core concept of the Spinal Implant EA. In this EA the algorithm takes a trade only if the fractal dimension is below a certain value.
-To apply walk forward testing.
This is a traditional approach and maybe is the state of the art in the trading community. But unfortunately quite a few algorithms pass this test, only to prove later that they are not so good.
-To apply Monte Carlo methods
So I really think that this should be implemented by everybody that seeks consistent results.
The Discussion will be here: only for logged users, sign up for free.
