Peer Review Trading System Research

[Rapunzel]

IMO neither of those are people you want to guide you in re to a quantitative strategy.

Researchers are not traders. They don't make money by trading and live in a hypothetical realm. You need a profitable quantitative/systematic trader to give a meaningful feedback. They will be from all walks of life as there is no school that teaches that.

I'm not volunteering just pointing out that you have wrong expectations. Based on what you mentioned - when results look too good to be true typically they are. Either problem with coding, data quality or slippage and cost assumptions.

Just trade it live with a small account. You will have good idea after 30-100 trades.

I have been working mostly with MT4 although I intend to move onto MT5 (I know, I know). The data is from TickData suite so the model quality on back-tests is usually 99%. The entry logic is simple and the idea generation process usually follows:

1. search the charts for market situations I want to trade
2. write code logic to attempt to describe those situations (Model Buys and Sells Separately)
3. test the code by taking pictures of open trades in a back test
4. revise 2 and 3 until satisfied

5. write a code with the above logic that also takes indicator readings and readings of my special filter, when trades open (multiple trade concurrent open trades permitted to capture every moment signal was valid). Print readings onto csv (yes, yes I know) with trade outcome 1 win or 0 lose.

6. Data Separation:
Entire Range will be 2004 - 2021
Model building range 2004 - 2017 (50% Model Build 25% and 25% validation)

7. model 'special filter' levels to find edge situations ultimately with simple linear regression where R sqrd > 0.95.

8. R:R on trades is 1:2 so breakeven win rate is 66.67%, searching for 77% win rate in model and validation sets with strict draw down conditions (1 loss in 3 maximum).

9. Test in period 2017 to 2021

10. Use the special filter ranges found and SVM classification to search for an edge in model and validation sets. (usually on predicting the losses).

11. Apply the SVM if something found so signals need unanimous vote of regression and SVM.

12. Test in 2017 - 2021

13. Bring online if okay and do the same for other pairs buy and sell separately.

I feel like it is quite thorough but there are a few tricks I might be missing. But as you said, the proof of the pudding will be in the outcomes from trading the signals live. At the moment with the signals and my final input I am doing 85% win rates, but; paper money.

 

[ValeryN]

I have been working mostly with MT4 although I intend to move onto MT5 (I know, I know). The data is from TickData suite so the model quality on back-tests is usually 99%. The entry logic is simple and the idea generation process usually follows:

1. search the charts for market situations I want to trade
2. write code logic to attempt to describe those situations (Model Buys and Sells Separately)
3. test the code by taking pictures of open trades in a back test
4. revise 2 and 3 until satisfied

5. write a code with the above logic that also takes indicator readings and readings of my special filter, when trades open (multiple trade concurrent open trades permitted to capture every moment signal was valid). Print readings onto csv (yes, yes I know) with trade outcome 1 win or 0 lose.

6. Data Separation:
Entire Range will be 2004 - 2021
Model building range 2004 - 2017 (50% Model Build 25% and 25% validation)

7. model 'special filter' levels to find edge situations ultimately with simple linear regression where R sqrd > 0.95.

8. R:R on trades is 1:2 so breakeven win rate is 66.67%, searching for 77% win rate in model and validation sets with strict draw down conditions (1 loss in 3 maximum).

9. Test in period 2017 to 2021

10. Use the special filter ranges found and SVM classification to search for an edge in model and validation sets. (usually on predicting the losses).

11. Apply the SVM if something found so signals need unanimous vote of regression and SVM.

12. Test in 2017 - 2021

13. Bring online if okay and do the same for other pairs buy and sell separately.

I feel like it is quite thorough but there are a few tricks I might be missing. But as you said, the proof of the pudding will be in the outcomes from trading the signals live. At the moment with the signals and my final input I am doing 85% win rates, but; paper money.

It does seem like you thought about this quite a bit. I don't see big reg flags. Couple of things you didn't mention that are quite important - number of trades and costs / slippage assumptions. I personally only used paper trading to test my custom execution software many years ago when I first created it and never use it for testing of individual strategies. There are many things that are not simulated there properly, including ones that might work in your favour in live trading.

Run it live on a small size, after 30 trades or so you will have a good idea on where to take it from there.

I personally think most people overthink their strategies too much (especially people with sciences/eng/math background). I've seen some spending a year developing something only to find out after a day of trading that it will never work because they can't get a fill or slippage is so terrible etc. So those experiments should be prioritized and done early to avoid a huge waste of time on unrealistic assumptions.

Also, size your positions based on hypothetical outliers, not just the ones you saw in the backtest.

Val

 

[Rapunzel]

It does seem like you thought about this quite a bit. I don't see big reg flags. Couple of things you didn't mention that are quite important - number of trades and costs / slippage assumptions. I personally only used paper trading to test my custom execution software many years ago when I first created it and never use it for testing of individual strategies. There are many things that are not simulated there properly, including ones that might work in your favour in live trading.

Run it live on a small size, after 30 trades or so you will have a good idea on where to take it from there.

I personally think most people overthink their strategies too much (especially people with sciences/eng/math background). I've seen some spending a year developing something only to find out after a day of trading that it will never work because they can't get a fill or slippage is so terrible etc. So those experiments should be prioritized and done early to avoid a huge waste of time on unrealistic assumptions.

Also, size your positions based on hypothetical outliers, not just the ones you saw in the backtest.

Val

Thank you again. Will go live with small risk post-haste. I suppose the model doesn't have to be academic perfection, so long as it works.

 

[easymon1]

Why not set a budget to limit costs of this research project, implement your strategy on a tiny tiny scale, generate some real world data and if you approach the bugetary limitations of your education budget, then put the strategy on hold and analyze the data you generated?

If it flies, buy me a beer.

So this is not a scholarly exercise or an attempt to obtain a position with an employer.

We all concur.
Let'r hap'n cap'n
Let'f rip skip.
Keep your eyes on your fries.

 

[Rapunzel]

So this is not a scholarly exercise or an attempt to obtain a position with an employer.

We all concur.
Let'r hap'n cap'n
Let'f rip skip.
Keep your eyes on your fries.

Lol. Nice rhymes. You're right about going live and letting that be the proof of the model.

 

[easymon1]

Lol. Nice rhymes. You're right about going live and letting that be the proof of the model.

You'll do fine. Break a Leg!