Losing algo-traders support group

[Energetic]

Sorry for the subject line - could not resist.



As someone here recommended, I've read "The Market Wizards" book. One of the things that I took away was that several wizards said that, for the purposes of risk control, they take the chips off the table when they lose certain percentage of money and then gradually return back to the market as their confidence restored.



I thought it's a reasonable thing to automate so I tried to apply several obvious ideas (i.e. reducing the bet size as the drawdown deepened, at various speeds) to my strategy ... and failed miserably. In the best version, the profitability and ratios declined only modestly while the max drawdown became a just bit shallower. In all other versions, I lost performance without improving drawdowns at all. The reason is that the strategy tends to recover fairly quickly on its own so by the time I was beginning to pull the money away it was typically just about to rebound.



Has anyone tried to do something like that? Any success?

[tabris]

"As someone here recommended, I've read "The Market Wizards" book. One of the things that I took away was that several wizards said that, for the purposes of risk control, they take the chips off the table when they lose certain percentage of money and then gradually return back to the market as their confidence restored."



My guess that this worked for them is that they have a decently high autocorrelation on their returns or their style/strategy has an inherent bias towards momentum.



"I thought it's a reasonable thing to automate so I tried to apply several obvious ideas (i.e. reducing the bet size as the drawdown deepened, at various speeds) to my strategy ... and failed miserably. In the best version, the profitability and ratios declined only modestly while the max drawdown became a just bit shallower. In all other versions, I lost performance without improving drawdowns at all. The reason is that the strategy tends to recover fairly quickly on its own so by the time I was beginning to pull the money away it was typically just about to rebound."



My guesses at why this did not work for you is... imagine you play blackjack where you have an edge. But you reduce your bet size when you lose money not when your edge goes lower. This will most likely result in what you described. So, the answer/solution to that would be, instead of looking at drawdown... you might want to reduce size on edge decay which is probably a harder measure. Second guess as to why this might happen is your returns might actually be negatively autocorrelated or your strategy trades some type of mean reversion. In this case, the optimal bet size is actually doubling down as edge has actually increased in your favor. By reducing bet size on drawdown, you are actually betting less when your edge has increased which will almost always end in ruin.

[london]

A few random musings:



Reducing risk after a run of poor performance might be a good thing or bad thing: depends on the strategy (or the trader!)



Good: when the strategy has stopped working; the world has changed; the edge you were capturing has decayed (either temporarily or permanently); then deploying less capital is good if the expected return is now negative.



Bad: if the edge is still there and just going through a temporary drawdown. The expected strategy return is still long run positive, but putting less capital behind each trade, it takes longer to dig out the hole.



So how do you know which is which? That’s the hard part!

Ongoing monitoring and having a view of when your strategy is not performing as expected.

1. Look at historical strategy attributes - per Tabris - autocorrelation; and also distribution of P&L per trade and per period, number of trades, etc, etc. Have a view on how different your live strategy attributes need to get before you think something smells funny. I’m a big fan of statistical process control and Deming. It’s not flashy or sexy but I want to know something might be different _before_ I’ve collected a long sequence of losses.

Its a small data problem- I want to know something is different before I have big data (and big losses and drawdowns).



2. Under what conditions has your strategy done well historically?; under what conditions has your strategy done poorly historically? And - critically - what conditions have not been observed historically?

I’m a big fan of knowing the range of conditions a strategy has been tested under to then understand where it has _not_ been tested. This is the true unknown and where to reduce the chips. Or at least pay closer attention.



3. Keep in mind you’ve never seen a strategy largest drawdown:a maximum (by definition) can only get bigger



4. The other time when cutting risk on losses is a _very_ good idea: for discretionary traders.

Have a fixed loss per trade, per hour, per day, per week or whatever. Hard stops. No more trades.

If one gets breached: close positions, stand up, walk away. Don’t come back till the next period.

When trades are being placed discretionary (rather than automated); this avoids the human over-confidence and emotional temptation of trying to “trade out” of losses with increasing larger and larger bets.

Eventually, that ends only one way and it’s not pretty.



I suspect discretionary trading is not your context.

But maybe this is the context of Market Wizards? I’ve never actually read it.





So if reducing risk on losses is good or bad, depends on: the strategy in question; how discretionary the trader is and the current environment versus the historical environment when the strategy was developed.

[ronin]

Yes, welcome to trading 101.



Stop losses don't help, and they will ruin any strategy. The only way to keep them from ruining your strategy is make the stops so wide that they are meaningless.



In the real world, you are running the strategy because it has certain risk characteristic. A drawdown may be just the cost of doing business, or it may mean that the strategy no longer has the risk characteristics you want. Or never had them in the first place.



That is what you analyse, and that will then drive your decision what do do with the strategy, not the depth of the drawdown per se.

[Energetic]

@ronin:



In other words, the wizards are actually rookies who didn't know what they're doing, never mind the mind-boggling performance results? Thanks, that helps a lot.



BTW, unlike (some of) them I didn't implement a stop-loss. My size functions were continuous.



Edit: Now it comes back to me that some of the wizards took an equally disdainful view on stop losses. But others have not. That was my other takeaway: there are many ways to skin the cat.

[Energetic]

@tabris, thanks - that's a reasonable diagnosis.

[Energetic]

@london



Thanks for taking time!



Of course, I am always concerned whether the edge is gone. One possible view on drawdowns is that the edge goes away temporarily and then comes back. Or it's just a string of bad luck. It does, by the end of the day, come down to autocorrelation which I doubt I can predict with any accuracy.



1. I do have my own version of ongoing monitoring. I described it [url=/Show%20Post.aspx?PostIDKey=188041]in this thread[/url]. General idea at the top of the thread and some details a little later in the comments.



2. Oh, I obsessed over this question a lot but I don't have the answer. Sometimes it looks like the signal temporarily disappears even though the market seems to be unremarkable. Actually, in terms of returns (not DDs) on a vol trading strategy a bad time is now. And I think I know the reason: someone is posting to Twitter too much



3. Yes, it's good line



4. Correct, I'm not discretionary. My "product" is an algo.

[gaj]

I think the pure quant approach would be Bayesian. You start by predicting the return distribution P(X). The sizing will naturally be a function of P(X). As new data comes in, automatically adjust the distribution to P(X | new data). This automatically readjusts your model if the new data supports or contradicts your hypothesis.



This requires modelling the full distribution P(X, data) beforehand though, which may not be so straightforward.

[gmetric_Flow]

Take a look at chapter 8 on Money Management in Filthy's book- you might find the section on Alternatives to Kelly fairly interesting. Alternatively, your favorite gambling book should discuss betting progression and of course the different schemes and how they fit into different utility curves.

[JTDerp]

(removed)