9
$\begingroup$

As I understand it, an algorithmic trading strategy could lose profitability, if, for example:

  • it's rediscovered by others
  • employee turnover leaks the strategy to others
  • market conditions change somehow

Is there a typical "half-life" of a strategy? Is it different for HFT vs non-HFT?

Improve this question
$\endgroup$
2
  • $\begingroup$ Yes, all three points (assuming it’s already profitable out of sample in live trading and doesn’t suffer from backtesting overfitting. Sometimes overfitted backtesting models can even work in live for 1-2 years before breaking down). I found employee turnover to be the fastest way to spread. A typically guy will interview with 5-15 other firms and each firm can implement the basic idea or a correlated variant. $\endgroup$
    – uday
    Commented Mar 19, 2021 at 1:20
  • $\begingroup$ HFT strategies can suffer from sudden abrupt changes in microstructure too, which non-HFT strategies are less likely to get affected by. Even simple changes like a Min Tick Size in future can render a market more momentum-y or more mean-reversion-y in the HFT space. $\endgroup$
    – uday
    Commented Mar 19, 2021 at 1:23

2 Answers 2

Reset to default
12
$\begingroup$

Is there a typical "half-life" of a strategy?

This is a really subjective question, and I don't think any singular answer will generalize well. That being said, I will give some examples from personal experience.

I have made hundreds of trading models in my career. I have only deployed 9 into live trading in the last ~25 years. Of those 9, 2 of them have been shelved indefinitely. Like what @SergeiRodionov wrote, those 2 models didn't slowly lose profitability; one stopped working due to changes in the way the NYSE disseminated data. The other stopped working due to order type changes on Nasdaq.

The other 7 have been trading continuously for very long periods of time. My oldest model was deployed in 2000 and still trades every day. My youngest model is 4 years old. I don't know the typical life span of a model because I have never deployed a model outside of my firm, so I really don't know if the ages of my models are atypical. I suspect that plenty of funds out there have long-term functioning models because I don't regard what we do as incredibly sophisticated. In fact, the oldest model is by far the most simplistic one.

Employee turnover has never affected our firm because all coding work is compartmentalized, and final assembly is done by one other partner and me, so only 2 of us actually have the proverbial keys to the castle.

Is it different for HFT vs non-HFT?

Yes. HFT is mechanical. If the way information arrives changes, the model can break. If the types of orders that your model uses change, the model can break. These types of breaks may be fixed easily or not--like my example above of having to shelve 2 models of my own. HFT relies on order flow and modern-day tape reading. If you read information faster than others and can act the quickest, you win. Everything else is 'non-HFT' as you call it, and that encompasses such a broad array of things that it isn't really worth comparing them.

We continue to try to develop new models and ideas constantly, but most never make it past out-of-sample testing. In my opinion, a few keys are:

  • Do not rush a model into deployment out of frustration or impatience. Most models will be thrown out. If rushed, as suggested in other comments and answers, a model could still work and even be profitable for a period of time; however, it will be anything but robust and will likely be a costly failure at some point.
  • A second key is to not over-complicate a model. The more parameters and moving parts, the harder it will be to optimize, tweak, fix, etc.

I could go on and on, but this is very subjective, as I already mentioned, and my answer is largely based on my opinions and experience, so I don't know how helpful it is, but I'm happy to add detail to anything should you find it useful.

Improve this answer
$\endgroup$
3
  • 3
    $\begingroup$ Can you share, in general terms, how profitable your working models are? For example, how long did it take before the models broke even, after all costs (including development costs) were taken into account? Do they pay for the cost of the electricity required to run them? $\endgroup$
    – Robert Harvey
    Commented Mar 19, 2021 at 18:12
  • 3
    $\begingroup$ @RobertHarvey The profitability of the individual models varies quite a bit, and they are all quite different in that they each perform better (or worse) in different environments. That being said, the most profitable on a gross basis (also most volatile by far) is my longest-term model (~21 years). It has a lifetime CAGR that is just over 49% with an annualized volatility of about 25%, so the high return does not come cheap, to say the least. My lowest returning model returns about 5% annually. However, its volatility is only 0.5% annually, so, as you can see, there is quite a difference. $\endgroup$
    – amdopt
    Commented Mar 19, 2021 at 20:07
  • 2
    $\begingroup$ Continued from my previous comment: The models are mixed and optimized in such a way as to attempt to produce consistent returns with the lowest amount of volatility that we can. Obviously, we do not make 49% per year--the volatility would cause us to lose all of our investors quite quickly for sure. It hard to measure the other things you are asking because I never kept track of that in my younger years. Electricity costs are de minimis compared to the amount of AUM the models trade though that would be something to examine if we were to lose significant assets. $\endgroup$
    – amdopt
    Commented Mar 19, 2021 at 20:13
5
$\begingroup$

6 months is a reasonable estimate. Typically such strategies do not decay as in half-time but rather stop working in a discrete manner. Parameter tweaking can help, but for a limited period of time. To add to your list which is valid in itself:

  1. Infrastructure changes at the trading venue, in particular latency-related changes
  2. Market-making program changes at the trading venue
  3. Order type changes at the trading venue
  4. Reference data changes (price increment, round lot size etc)
  5. Listings/de-listings of correlated instruments
Improve this answer
$\endgroup$
1
  • $\begingroup$ Given no comission, no leverage, no side costs (electricity etc.) and financial instrument existing infinitely long, market making is forever profitable in financial terms, i.e. generates profit over instrument price. Sadly none of those conditions ever true at once. Market makers are not going to risk their money, they are only risking bankers money. $\endgroup$
    – sanaris
    Commented Mar 19, 2021 at 21:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.