I’m not sure if there’s an official name for these types of trading strategies so I’ll call them “Trigger Based”.
Basically, they have at least 1 rule of the following type:
If X > A, do something (i.e. buy/sell the stock), where X could be the stock price, an indicator (i.e. 50 day moving average), etc. The number “A” is the trigger value… could be 0, 100, or whatever.
The problem is if X narrowly misses A, then no action is taken… it’s all or nothing.
My question: is there any name for this sort of class of trading strategies or quantitative models? I see them everywhere these days (suggested by amateur traders as well as published in finance magazines and journals.) Are there any academic studies that prove/disprove their effectiveness?
We trade a system like that. It is frustrating. On the other hand, you don’t do it unless your methodology has been appropriately backtestested and you understand the assumptions you are making when you implement a rule like that. You realize that you will sometimes have “near misses” and then you’ll start to get extra sensitive about whether the near miss resulted in a net positive or negative decision.
There are ways of smoothing this, so that you start to build a position at A-delta and get to a full position at A+delta, but whatever delta is is usually an arbitrarily chosen number.
Backtests of course are not everything, but they are still important. A set of rules that doesn’t perform well in a backtest is basically a non-starter unless there is a very compelling argument for what has changed that will make the system perform better in the future.
Thus, backtesting well is (most of the time) a necessary condition for a rules-based system to be a good candidate, it is not a sufficient condition, though.
The solution to this is to run backtests at different thresholds (i.e If X>(A-1)) to find out what the signal bounds should be. What I’ve found is the cases where X narrowly misses A aren’t usually important. Meaning when those situations arise it’s not a ‘strong’ signal and the outcomes are basically noise - you can then choose to omit these ‘narrow miss’ signals or modify the rules further (or trade them anyway).
More importantly though, any good system will work even if the signals aren’t traded very well. Meaning success won’t be dependent on X being narrowly bigger than A so a signal can be triggered. Often times you can ignore signals or not trade them all, and if it’s a good strategy it’ll work anyway. For example, over the past few years any bullish strategy would’ve done well and the specific parameters wouldn’t have mattered because the market just went up. Don’t trade anything that requires perfect execution to succeed, that means the margins are too small and you should look elsewhere.
As far as non ‘trigger-based’ systems - what else is there? Presumably there’s a reason, signal, market condition, or event that makes a trade look reasonable. The only alternative is an always-in strategy, no?
One main advantage is that these types of systems are very computationally efficient. This makes them more likely to be used at very high frequencies. However, most (including me) have no need for that.
Anyway, I’ve built a decent number of these types strategies. When I first started developing them, I thought they were the bee’s knees, but I’ve come to more-or-less hate the approach. Mainly because it’s far too easy to convince yourself that something is really hot shit by adjusting the parameters when it will actually perform terribly out-of-sample. Bchad’s approach partially ameliorates the binary nature of these strategies, but it doesn’t really resolve the issue.
I would go further and say that academic finance relies too much on these as well (e.g. go long/short the top/borttom 20% book to market stocks, rebalance in a year).
@LPoulin133 I do not think that “always-in” is the only alternative. Look at what Bchad wrote about scaling, for starters. Alternately, for your issue about bullish signals working, you could evaluate individual stock signals based on the residuals from a regression of the stock returns on market returns. Finally, portfolio optimization (and what practioners have done to improve on it) can be used to generate trades and account for transaction costs.
My point about scaling was that if what you hate is the on-off nature of these strategies, there are scaling methods to reduce that particular frustration.
Various portfolio optimization techniques can also serve as a blending tool so that nothing is quite so on/off as the raw trading rules would imply, as jmh pointed out.
I agree that the bigger problem is that these rules make it extremely easy to overspecify things to make them look great in backtests, but then perform terribly going forward. These methods are extremely easy for the unscrupulous marketer or the dim manager to glom onto.
There are some methods to try to deal with that, such as separate calibration and testing samples, monte carlo methods, etc… None of them are bullet proof, but it is still important to use them where feasible.
Investors have always had trigger-based trading strategies.
The only thing that the computers have done is to broaden the criteria in the triggers, made the triggers more precise (and maybe more accurate), made the trades quicker, and given more idiots the ability to trade.
Fundamentally, nothing’s changed. Except maybe the idiot part.