signal vs pnl

[nikol]

There are periods when trading Signal consistently delivers Profits and periods when there are serial Losses. I notice that if I bluntly reverse the sign of the Signal the system delivers Profits.



Is it a smart idea to introduce sign = {-1,1} which depends on correlation (Signal, Profit/Loss)?



PS. Periods are measured in days, although I might trade on 15-60 minutes frequency.

[ronin]

In other words, your signal is more autocorrelated than the market you are trading.



It is a smart idea to work on breaking down the autocorrelation.



But the simple hack you are proposing will in all likelihood not work. That sort of blunt instrument never does. You will probably need to dig deeper into where the excess autocorrelation is coming from, and redesign the model around that.



It is also possible that the autocorrelation is your signal - once you remove it, there is nothing left.

[doomanx]

There's a parallel here with the idea of just looking at a bad trader and doing the exact opposite of what they do. The opposite of random is random, so if you don't understand why it's generating some P&L and can't reasonably explain the variation, I would recommend against it. Note that this idea isn't useless (otherwise b-bookers wouldn't exist), but it's not as simple as just 'do the opposite'.

In particular want to highlight the need to make sure you're not just overfitting by flipping your trades when they're losers and making them into winners - using this method you could make any strategy in the world into a goldmine on a backtest, but there's a big data mining risk. As Ronin mentioned the autocorrelation might be the sauce and without it you just have noise. Proceed cautiously.

[nikol]

@ronin



"You will probably need to dig deeper into where the excess autocorrelation is coming from, and redesign the model around that."



I was afraid about this. Thanks for confirming.



But do I have the autocorrelation really if my trades are generated by analysis of order_book at tick level with ~15-60 minutes frequency (10-50 trades/day), while I look into series of resulting daily PnLs (=aggregated per day) and would like to trigger after observing N-days it is negative?



Or you are talking about PnL-autocorrelation which might indicate, for example, some signal at "macro"-level?





@doomanx



"big data mining risk"



True.

[ronin]

@nikol,



I assume you do have autocorrelation based on the wording of your post, but I don't know. Run some autocorrelations on your data and see what comes out.



I am talking about pnl autocorrelation, yes.



A simple explanation would be that your pnl is linked to the volatility of the underlying. You make money when the volatility is in some range, you lose money when it isn't. Volatility tends to cluster in ranges, so you get strings of winning days and strings of losing days.



But, at the end of the day, I have no idea.

[nikol]

"A simple explanation would be that your pnl is linked to the volatility of the underlying"



Of course! Best example, MM strategies with most wins at low volatility + high trade volume vs Momentum stats where high volatility is preferred.





This all was useful. Thank you, gentlemen.

[Energetic]

@nikol



How long are your losing sequences?



I looked at my strategy and found one incident of 7 consecutive losses and a few of 6 over 14 years. Tried to flip sign after 3,4,5 consecutive but it didn't do me any good.



I did't think it's a good idea even before I tried. I believe any strategy looks only at a certain sub-space of the actual state of the world. Sometimes, there is not much happening in your sub-space vs. what is going on in the orthogonal. (Maybe it's your chance to revisit the strategy and figure out what you are missing?) Ideally it shouldn't happen but it does. And then there is genuine randomness that cannot be captured by any strategy. Sometimes you get tails 10 times in a row. Obviously, you can't deal with it in any rational manner.



Apologies if I sound as if I know what I'm talking about. These are just my views that may be completely wrong.

[nikol]

@Energetic



Periods are something like 30 days in profit, 10 days in loss. And it does not look as a random walk.



Thank you for trying, but I came to conclusion that @ronin was 100% right - I missed parameter.

Very likely it is volatility. And, also very likely, it is difference btw Difficulty and Hash Rate )

[Energetic]

Missing parameter is essentially what I meant when I mentioned orthogonal sub-space.