Transactional vs Confidence-based Trading Strategies
Warning: extreme geekiness ahead.
I think folks sometimes have a difficult time understanding my approach to the markets seen in YK, Scotty, and the State of the Market report, because most traders come from what I’ll call a “transactional” school of thought rather than a “confidence-based” one.
Note: I just completely made those words up, so don’t bother googling them.
A traditional transactional strategy might go something like: buy X shares when moving average X crosses over moving average Y – sell the position and go short when it crosses under. I call this “transactional” because if we were to begin trading that strategy after a buy signal but before a sell signal, the strategy doesn’t work as designed because it relies on the sequencing of the market, or put another way, starting at the beginning of the trade and finishing at the end.
A confidence-based approach to that same strategy might look more like: as moving average X climbs farther above/below moving average Y, increase exposure to the market from -100 to 100% (and all the steps in between) with 0% being X=Y and +/-100% being X = +/- 3 standard deviations above/below Y.
That’s a little difficult to explain in text, but the difference is two-fold: (a) you could jump into the strategy at any point and it would have an opinion on the market, and (b) the strategy isn’t just trading a condition (i.e. X crossing Y) but expressing a confidence in that prediction from -100 to 100% and all the steps in between.
Side note: obviously you could scale into a transactional strategy as well and also call this a “confidence in the prediction”, but the vast majority of the time, transactional strategies do not – they have very binary entry/exit rules.
The confidence-based approach has two advantages.
First, and most importantly, it increases sample size.
In the transactional example above, each closed trade is a single observation. For something like 50/200-day moving average crossovers or a small/large-cap leader/laggard strategy, it could take multiple decades to generate enough trades to have any level of confidence in the strategy.
But in the confidence-based strategy, because of the introduction of the concept of confidence (and the fact that it’s radically changing from 0 to 100%) smaller units of time (such as days) can be treated as an observation.
Side note: well, not completely…in a very long-term strategy like 50/200-day crossovers the confidence value isn’t changing enough day-to-day, so no, each day is not an observation – but perhaps each month is, and that’s a hell of an improvement.
The second advantage is that it forces the trader to think of the portfolio as a sliding scale rather than X number of fixed positions (again, this could also be accomplished in a transactional system but usually is not). Not only are we using the binary condition to guide our trades (X crossing Y), but the strength of that condition as well. That added layer can do a lot to help a strategy focus exposure on those times when the market becomes particularly predictable and reduce exposure when it is less so.
The Kicker
The fly in the ointment is that most traders don’t have a sufficiently large portfolio to employ a confidence-based approach that is making fine adjustments to position sizes on a daily basis. Transaction costs would kill the goose.
That’s a big part of the reason why I trade leveraged mutual funds. No per-transaction costs mean that my own portfolios and all of the investor portfolios that are following our strategies via managed accounts receive, for all intents and purposes, the same results, regardless of the size of the portfolio, even though we’re changing positions (sometimes only very slightly) everyday.
Just one humble developer’s $0.02.
Happy Trading,
ms
P.S. this will be the post I will point folks to when they ask me when the Scotty or YK strategies issue a sell signal. The answer? They never issue a sell signal…for that matter they never issue a buy signal. Every day is just another change in the strategy’s confidence in a given direction.
To stay up to date with what’s happening at the MarketSci Blog, we recommend subscribing to our RSS Feed or Email Feed.
Filed under: Trading Strategies | 14 Comments
Very good stuff Michael as usual – and it gives me some insight into the tools you use. I can’t think of a single trading platform where expressing a system in terms of confidence is particularly easy. Now, you trade a very limited number of instruments – so the data handling aspects of Excel don’t matter much – you’re not dealing with, say, 1500 stocks each with 10 years of data. But what Excel would clearly be good at is expressing a confidence-based system (by the way, with Madoff out there, that name is probably not a good one :) ). One thing I like about your approach to research is this concentration on a small number of instruments – I think it helps one focus. One question – so with YK, for example, the confidence factor must switch quite often? I’m curious what it means in the context of a fast-changing system.
RE to dskills: in all three of the places where the concept is public (YK, Scotty, and the SOTM) the confidence is usually changing pretty drastically day to day even if the direction of the prediction isn’t. I think that’s important to the “increased sample size” argument.
RE: the right tool – you’re right, Excel wouldn’t be appropriate for a large db of stocks. I wrote a post a while ago about the tools I use (http://marketsci.wordpress.com/2009/01/06/faq-software-and-data/) and for this particular challenge I would definitely use Perl or some other programming language…just too customized I think to be done in any sort of off the shelf program.
Long time no talk – it was good to hear from you…
michael
It seems to me that you’ve only gone half-way with this explanation. Another advantage to confidence-based investing is that it has better noise rejection. For example, if you have a transactional system that says to buy when RSI(2) drops below 10, it says to do nothing if RSI(2) drops to 10.01 and to go fully long when RSI(2) drops to 9.99, even though the difference is generally less than a penny in the stock price. With a confidence based system, the rule might be that at RSI(2) of 10.01 you should buy 1 unit for every $900 and at RSI(2) of 9.99 you buy 1 unit for ever $895 dollars, which is only a small difference that better reflects the lack of any meaningful difference between the two indicator values. With moving average crossovers, you might be able to come up with rules that allow you to avoid whipsaws with confidence-based methods, which is impossible to do with transactional methods. But whipsaws are an inherent part of momentum methods, so they probably can’t be avoided entirely.
Furthermore, confidence-based approaches allow you to combine multiple trading strategies more easily. With transactional strategies, there is no obvious way to combine two systems to make a better one (the easy way to combine them is to allocate some fraction of your capital to each system). With confidence-based strategies, you have confidence numbers that can be combined in accordance with statistics to produce a better confidence estimate (as is done in the state of the market reports).
RE to jkw: well done sir. I completely agree with everything you’ve written. michael
Thanks for the post ,what a concept!
Mike – can you comment on how to develop a confidence based system? I’m trying to come up with a simple example.
Sure…give me a week or so. I need to do all of my end of month stuff this week. michael
jkw- good point. That is also important.
I think instead of “transactional”, perhaps a more accurate term would be “binary”- either they’re in a “trade” mode or a “non-trade” mode.
I agree with you confidence based systems are more interesting, accurate, and the various other advantages. But they’re also much more difficult to develop and tweak- another one of the reasons why most systems are still “transactional”
Speaking of confidence, I’m playing around with a ‘system’ that buys market weakness in small increments (up to a very conservative % total of my capital) and uses your ‘state’ daily aggregate prediction to help guide me on buying points (among other general indicators). Throwing good money after bad from time to time with no stops :) but managing risk via using indexes only and small position sizing on individual buys and in total. Selling on the pops back up.
I’m going to keep trading that way until ‘things get better’ (or death).
You can also have a hybrid type system, which I would describe as one that has some kind of transactional entry/exit signals but uses confidence (or expected value) to determine the size of the position.
One example is where the system says buy & you know (based on past results) that past buy signals with the same traits (Positive Slope, High Volatility & High Volume for example) as the current one generate a very large & consistent profit, so you increase your size accordingly. Conversely, a sell signal with Positive Slope & Low Volatility may get a smaller position size.
Breaking your systems down into the conditions present when the trade signal was generated can yield some very large benefits.
Eric
Michael,
I agree with you after the following is accounted for:
1. first and foremost, you have to makes sure the variable and it’s relationship to E[r] is normal if you’re using a normal distribution, which I doubt it is.
2. you need to account for the dynamics of this relationship over time.
3. Of course the stability of the model
4. stability of the model components
5 All of this over sample data.
6. when you’re doing large scale trading, you have to come up with a (cost/benefit relationship to this type of scaling) vs. (the efficiency of transaction and information you’re giving the market)
I would say that most people do not account for many if not all of these and would fall victim to the “i’ll keep using these paramters or set of paramters till something breaks”
After all that you then have to realize that our models still aren’t perfect and that risk control should cover this.
my .02
Could you explain how you calculate the standard deviation of one MA above/below another?
Thanks.