Why does a given model expire

 

Everyone says, especially from manual trading, so as not to complicate trading (eat, sleep, trade) :) it is enough to apply simple rules: trend, eg EMA, price action, trigger, sl and tp.

Repeat this a great number of times and the profit will come. I try to check it somehow, and so far I have no confirmation for this approach, or I do not understand something.


Suppose we have a simple trading model:

1. H1
2. TEMA as trend and trigger
3. Signal and trigger are cross TEMA with Pin Bar type candle (TEMA rejection)
4. Stop Loss and Take Profit
Suppose the EA code has no technical errors and that the losses are not due to errors in the code.

I optimize the model in terms of three parameters of freedom: the TEMA period and the number of SL and TP points for the EU. I am trying to generalize the model.

I optimize on the training set and test on the test set.

The model works ok with this data and some data out of range.

1. Repeating this n times almost always causes a loss
2. I have a reflection that in each period there are some specific values ​​of these three parameters of freedom for which the model is profitable.
3. Theoretically, each model can be profitable, but for some reason the parameters of freedom are also some variable, a function of something ...

What actually causes the freedom parameters to become obsolete in trading?

Possibly what I do or think wrong, maybe e.g. I wrongly assume that the trading model can be so simple ...

What are your thoughts?

PS. I will add that I know a bit about modeling and AI and understand the issues of overoptimization and adjusting models.
 

I did some manual "Walk-forward" tests.

I reduced the number of optimization parameters to two: TEMA Period, TEMA growth factor.

Model still is the same: H1, TEMA, trigger rejection TEMA by Pin Bar, fixed SL, TP and lots.

Results of individual "walks" 12 months for optimization, and 2 months for forward tests: 


I assume that for subsequent "periods" both parameters of freedom will constantly change their optimal values.

and I wonder what the results say:

1. In general I understand that "walk-forward" is really looking for the optimal period in the past for which the prediction of a certain period in the future is best.

Hence, I conclude that apart from optimizing the parameters of freedom, I should "look" for the optimal period for optimization vs the period for prediction.

2. Another thing, from what I can see, TEMA period and Growth Factor are probably not significant features, because basically they are different in each period.

Does this mean I should look for other significant features? can such results already be somehow meaningful to interpret?


What do you think and how do you see it?


For comparison, I will do two tests:

1. The same two parameters of freedom, but I will change the model "learning" period to 6 months and the prediction period to 1 month

2. I will add additional SL and TP freedom parameters for optimization.

 
Sebastian Rafal Skrzynecki #:

I did some manual "Walk-forward" tests.

I reduced the number of optimization parameters to two: TEMA Period, TEMA growth factor.

Model still is the same: H1, TEMA, trigger rejection TEMA by Pin Bar, fixed SL, TP and lots.

Results of individual "walks" 12 months for optimization, and 2 months for forward tests: 


I assume that for subsequent "periods" both parameters of freedom will constantly change their optimal values.

and I wonder what the results say:

1. In general I understand that "walk-forward" is really looking for the optimal period in the past for which the prediction of a certain period in the future is best.

Hence, I conclude that apart from optimizing the parameters of freedom, I should "look" for the optimal period for optimization vs the period for prediction.

2. Another thing, from what I can see, TEMA period and Growth Factor are probably not significant features, because basically they are different in each period.

Does this mean I should look for other significant features? can such results already be somehow meaningful to interpret?


What do you think and how do you see it?


For comparison, I will do two tests:

1. The same two parameters of freedom, but I will change the model "learning" period to 6 months and the prediction period to 1 month

2. I will add additional SL and TP freedom parameters for optimization.

Robert Pardo says there are 2 important parameters more for a strategy: number of Walk Forward stages and the IS:OOS ratio. You have to find them with Walk Forward Matrix. 

If you want to talk about it, I am here for it.

good luck..

Reason: