Kelly criterion

Beginner Questions
21

My understandng of this is that one of the many problems is that your edge isnt constant, it fluctuates, and you dont know what it is, you only have an estimation based on historical data.

To simplify things, lets assume on average its 5%, but in reality, there’ll probably be a short term distribution between maybe -5% and +15% (unless of course you are lucky enough to be one of the many traders here at babypips who makes a thousand pips a week and never loses)

So at times, your edge might be performing up at +15%, but using kelly, you’ll be sizing the bets as if you where at +5%, therefore you’ll be underbetting. So whilst its true that you may be making money, and compounding nicely, the gains have less marginal utility because you would have done well anyway during a winning streak.

Conversely when your edge is at -5%, you’ll be overbetting, and in extremely dangerous territory, and possibly even betting at a level greter than optiml F. This is the real problem with the approach.

I’m not really sure what else I can share, I have a mountain of backtest results, and they show what you’d expect too see, massive volatility, and all or nothing style returns.

I know that people use average win and loss, and they track ongoing win rate etc but at the end of the day, they’re estimating future perfomance based on history, and whilst the theoretical framework is sound, you cant actually quantify your edge accurately enough to exploit it.

I’m not saying that it doent have a place, the returns can be spectacular, but its probably not something you should use to trade your total net worth !

22

LOL - but every site is full of winners isn’t it? :slight_smile:
Particularly the ones selling videos that are actually set up as businesses :slight_smile:

So at times, your edge might be performing up at +15%, but using kelly, you’ll be sizing the bets as if you where at +5%, therefore you’ll be underbetting. So whilst its true that you may be making money, and compounding nicely, the gains have less marginal utility because you would have done well anyway during a winning streak.

Conversely when your edge is at -5%, you’ll be overbetting, and in extremely dangerous territory, and possibly even betting at a level greter than optiml F. This is the real problem with the approach.

I’m not really sure what else I can share, I have a mountain of backtest results, and they show what you’d expect too see, massive volatility, and all or nothing style returns.

I know that people use average win and loss, and they track ongoing win rate etc but at the end of the day, they’re estimating future perfomance based on history, and whilst the theoretical framework is sound, you cant actually quantify your edge accurately enough to exploit it.

I’m not saying that it doent have a place, the returns can be spectacular, but its probably not something you should use to trade your total net worth !

The fixed finance model must suffer from something similar though? Even compounding up your account can have the same issues. I am of the opinion that compounding daily is too volatile as if you win 1 trade and increase to only lose the next, you have lost more than you started with. I am comfortable compounding weekly due to weekly targets but it presumably has the same problem comparing week on week.

23

Hey SanMiguel, how’s it going?

Magnus and I came across the Kelly Criterion a couple of months ago and I did some playing with a spreadsheet. 25 columns of 100 trades to give an average expectancy, so I can say with this with certainty.

The Kelly Criterion works. I’ve found that for quite a range of win ratios and RRs that 25% works surprisingly well.

Why the discrepancy? Well I tried trading at 25% and lost, quite badly. The reasons were two fold - firstly the method I’d chosen to test this on (with real money as usual, I’ll never learn on that count) didn’t have the win-rate I had initially predicted - as you say you need a good sample size to be sure and I didn’t do sufficient due diligence.

But the main reason the Kelly criterion is so hard to trade is the root of all things trading - psychology. You can trade 1,000 into 10,000. Maybe even 10K into 100K. But not many people can place a trade with 25K risk, knowing that’s pretty much a year’s wage to be won or lost on the next swing of the market.

Edit: Here’s the thread where we touched on it before and the related theory of infinite yield. See http://forums.babypips.com/forextown/31062-dead-pips-2.html#post158674 and four or five posts following it.

24

I don’t think one should up their risk by taking on more positions.

The increase in gains, can come in many different forms.

Instead of increasing lots, check your losses. If you had added the amount of risk to your stop loss, would that have made the difference on a sufficient number of trades to more than cover the fewer bigger losses?

Even if it means getting several trades back to b/e, it should make a significant difference.

In my case it did considerably.
It would have upped the win percentage noticeably as well.

25

It’s not so much the upping of risk for the sake it but how a trader defines their optimal trade size. For example, yo could trade with 1.5% risk and compound every week - it works well for me. However, is that the optimal trade size for my strategies?
Strategies is a keyword there as well because no matter whether I average my last 50 trades, I use a slightly different entry criteria on each, which Kelly is not designed for. However, had I 500 trades worth of data, then I might feel more comfortable using it.

It is worth noting at this stage that 10% is a very standard investment amount for stock traders in the US owning equities where they probably wont lose the whole 10% of their trade.

However, no-one ever discusses how we as traders using margin/leverage come to the 2-5% amount either - is it a mathematical model or just accepted trader chat…
People talk about the worry of a losing streak or risk of ruin but that in itself is dependent upon the individual trader.

26

Good points San Miguel.

Trying to gauge results based on several strategies could definitely cause skewed percentages.

I think ultimately, using the Kelly, it comes down to risk management, and risk tolerance. Well, that and using one consistent setup for trade entry to gauge your numbers by.

And I also think risk is where the 2-5% “rules” came into play as well.

Since forex is leveraged so highly, it would be only logical to risk less than the counterparts in equities.

27

Well, I’m a bit late to the party on this one but I’ve always viewed every bet I make in terms of Kellies and have been struggling with how best to calculate edge (Estimated value/EV) especially in the EUR/USD on an hourly chart since I’m concerned that the sample size is a bit too small to be statistically valuable. Any suggestions on how best to get this done without sitting at my desk and blowing off my day job for the next few weeks?

As someone who has used Kelly for the better part of 6 years now in the blackjack arena I can tell you that it works but you have to understand it inside and out. I tried to post a link to the original paper but I can’t yet- not enough posts. I’ll try to throw it up later if anyone wants is.

The biggest problem I’ve seen using the criterion is not that it’s not sound- it is - it’s that people don’t use a large enough sample size and make their judgment based on statistical noise.

I’ve found you have to view kelly in a law of large numbers framework. My favorite example is a story from the MIT blackjack kids (who incidentally use kelly to determine their bet sizes based on count and other “games”). One of the big players lost $250,000 in three consecutive hands, but as far as the team was concerned they made money because the bet was properly sized and made at the correct/most opportune time. They lost in that single instance, but kelly and EV isn’t about a single shot even though we as traders and gamblers have a tendency to view it that way.

Anyway, sorry for the tangent, just hope it helps.

28

I also think so.

29

I’ve always been interested in a KC’s MM. I’ve been thinking for a few months now to open up a micro account and drop a few bucks in there so I can test it out with the system I use. I have a descent sample size but I don’t think I could personally stomach all that risk off the account I live off of.

31

Have you demoed?

33

Of course.

34

Lol actually right after I made that post I said to myself “why don’t I just demo it?” So I’m going for it now, I’ll see where I stand come December.