A Question for the fellow Quants
For all of the quants on here, what is the approximate proportion of strategies/factors that turn out to be useless to strategies that are successful? I'd also love to hear from anybody that has experience with FX algos (as i'm limited to only trading FX in my personal account since I'm interning at a hedge fund)
Useless as in the strategy or factor does not work out of sample, or useless as in they are not allocated any risk?
useless as in the factor does not work out of sample
most of time they are all garbage
I’ve traded crypto if that counts, looking at structure of order books and arbitrage
On your own account? Is it possible to make money this way?
Genuinely interested
I’ve heard crypto is expensive to trade. I can’t believe there is a whole lot of arb left since a lot of quants flocked to that space
99 out of 100 strategy ideas i've had were worthless
it depends on a variety of factors, but i'd say definitely a lot less than half as a gross generalization
Wait, so to clarify, over half of the time you test out a factor that actually works ?
no, a lot less than half the time a factor works
Even a lot of published asset pricing literature is questionable. Negative/inconclusive results are never published so there's an incentive play around with the data until it satisfies your original hypothesis.
It's hard question to answer. Unless your strategy is pretty high frequency, you won't know if your factor is useful for potentially years. If you're rebalancing the portfolio bimonthly, you're generating only 24 data points per year. To get a strong t-stat on whether long/short return on the factor is positive, you might need two or more years of data.
intra-day quants can get more immediate feedback on the efficacy of a factor or strategy change but most quants have to live with not really knowing whether what they did was useful out of sample for a long time. I don't know about higher frequency factors as that's not my space but most "useful" factors don't have that high of a sharpe in backtests. 0.4-0.6 is very common for simple single factor long/short sharpe. Most factors in backtests will go through multi year periods where the return is negative. But as long as the return profile is somewhat stable over time and the alpha hasn't decayed over time and is a diversifying source of alpha then it's okay to add.
Maybe higher freq quants can chime in on this as my experience is really mostly with monthly time series data.
In intraday market-neutral Sharpes are much higher (>2 generally).
As in a single factor in an intra-day strategy would show a sharpe of 2+ in a backtest?
Very interesting. Couldn’t you just go back further in time for the in sample data, freeing up more out of sample data?
By "out of sample" I mean actual post implementation performance, not in terms of model backtesting
You can. However, that is assuming the data quality and relevance holds up, and maybe also that the effect is not arbitraged away.
[double post]
Only about 10 - 20% of the time for lower frequency strategies. I think a key caveat here is how we define if something is "working" out of sample, which probably varies across category of strategies and from person to person.
For intraday strategies the success rate is probably higher, but only without trading assumptions. Optimizing the latter is arguably more important.
for >1d holding periods 10-20% sounds about right but this assumes you are not already adding to a book. In which case if you look at correlations you probably end up even lower as pre-existing strategies usually have overlaps.
Tcost management is critical and probably 1/3 of your realized sharpe for intraday. I'm assuming that's what you meant by trading assumptions.
Yes, disregarding correlations. Just purely talking about the number of ideas which eventually work out.
For intraday, it is my belief that there aren't actually that many different sources of alphas. So the variances in performance between different teams are largely due to how well they execute. T-cost, impact, participation rates broadly speaking. That is what I meant by trading assumptions.
I was going to say, if you have a bunch of 2+ sharpe factors that aren't correlated, the combination of many 2+ sharpe factors could produce a 5+ sharpe overall strategy, which seems unrealistic even for intraday. But maybe I'm just unaware
Dicta qui quidem beatae rem et in aut. Cupiditate similique commodi qui sequi qui quo consequatur eveniet. Ex aut vel amet architecto consectetur debitis dolore.
Officiis cumque aut ut maxime. Ipsum rerum corrupti totam eaque voluptas. Aperiam eos veritatis quaerat qui magnam quia.
See All Comments - 100% Free
WSO depends on everyone being able to pitch in when they know something. Unlock with your email and get bonus: 6 financial modeling lessons free ($199 value)
or Unlock with your social account...
Itaque quia sed autem quam. Sequi molestiae nobis aut dicta. Distinctio voluptas officiis est corrupti. Nisi laboriosam minus eum reprehenderit. Eos totam maxime officiis delectus fugit blanditiis temporibus sit.
Consequatur non deleniti maiores quasi veniam quaerat voluptatibus. Excepturi laborum atque pariatur dicta. Nulla id rerum nemo quia aut dolorem incidunt.