Here is the Pyfolio Analysis of the above post...FYI...
alan

Factors just not good enough/too much noise?

@Albert, nice factor!
Guess the question is still the same for either...what caused mine to drastically go down at the Pandemic Event...what caused yours to drastically go up.
alan

@Alan Thanks! I'm using an alternative data source which is extremely predictive even in times of high volatility.
Does your factor combination have an economic theory behind it?

@Albert ,
The strategy's economic thesis is that the sum of ~23 smoothed factors combining the best of size, value, volatility, momentum, and mean reversal factors will result in factor that captures the overall market tilt , then buying 200 long and 200 short, based on their aggregate factor value is supposed to move slowly, and be immune to market movements.

In looking at the "Cumulative Common Style Attribution" chart, I can see the "short_term_reversal" trace going crazy during the event.
My current working thesis is that either the shorts didn't cover in time(no market) or the shorts lost money simultaneously with the longs...somehow...

With respect to your graph, if the market did not obey basic protocol, why did your returns out-perform during the event?
Still a mystery to me, yet fun to talk about and figure out how to put in a Risk-On/Risk-Off Regime factor for a strategy.
alan

In addition to Albert's factors 'preferring' volatility, and Alan's factors don't, another reason might be that Albert's factors are more 'unique' and Alan's factors are more common. Let's say that a lot of funds have positions similar to what Alan's strategy holds, and in the last month when everyone deleverages (risk-off) at the same time, that means commonly held longs are being sold en masse, and commonly held shorts are being bought back. In other words, the spread for commonly implemented quant factors are moving in the wrong (wider) direction.

Quant Quake 2.0. I believe we saw an early ripple with Momentum stocks in Sep 2019.

@Joakim Exactly. My goal was to work with datasets not used by anyone (yet).

@Joakim,
Thanks for the thesis,,,actually sounds feasible...I'll create a cumulative returns on Shorts Only and Longs Only, and will report back here!
Guess I should also look at what happened to OrderFillPercentage during that month. Also, thanks for the video interview you did with Q...was instructive!

You used the word 'deleverage' above, and while I've heard in these forums that a neutral long/short equity strategy would be 'levered-up', does that practically mean buying Kx the shorts and Kx the longs, hence increasing the risk of cleanly "getting out" if your tear sheet doesn't go according to plan, like with the Pandemic Event.

Also, I believe that the average of the factors I've used have a response frequency much lower than the Event response frequency, which is a problem in that even though I'm rebalancing daily, the factor response is at an implicit weekly/monthly time frame. which means that even though the tearsheet shows everything as neutral, it's response time frame is way off, giving rise to no self-healing feedback, even with daily rebalancing.

@Albert, So by using unusual data, you escape the cascading "deleveraging" of common assets that @Joakim describes, right?
I just recently attended a webinar by @DePrado, and he mentioned "NowCasting" instead of "ForeCasting", essentially by scaping millions of scraps of data and machine learning what the current "state" is of various regimes and measures. THink this link was in another Forum post, yet here it is again:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3562025

I can't figure out from your chart why the Event made yours rise so much, yet I suppose that is rightfully proprietary!

alan

Hi Alan,

Regarding this:

You used the word 'deleverage' above, and while I've heard in these forums that a neutral long/short equity strategy would be 'levered-up', does that practically mean buying Kx the shorts and Kx the longs, hence increasing the risk of cleanly "getting out" if your tear sheet doesn't go according to plan, like with the Pandemic Event.

I don't quite understand what you mean unfortunately, but I'll try to explain what 'leverage' means. If a hedge fund has $10mil of investors money deposited with their broker, depending on the type of strategy the broker might offer them a credit line to increase the total capital that can be allocated to that strategy. Let's say it's 10x of the current value of cash/securities deposited. So now the fund can allocate $100mil to the 'risk neutral' strategy (with certain risk constraints/parameters that need to be monitored and within bounds - e.g. VAR for example, and maintaining a minimum net liq of say 1mil). 10mil of investors capital + 90mil credit from the broker. Deleveraging just means reducing the amount of leverage applied, so going from 10x leverage to 5x leverage, or down to 50mil capital deployed in this case. If everyone does the same thing at the same time, it will move commonly held positions in the 'wrong' direction.

Note that VAR has it's flaws, and it's not very good at accounting for black-swan events (LTCM, GFC, Covid-19) so it's probably not used much anymore, favouring other risk models

If a fund is in 'violation' of the broker's risk model, the fund will get a 'margin call' from the broker's risk team. The fund will then either have to top up with some fresh cash/securities, or trade out of the positions contributing most to the violation (or some combination of this), in order to bring the fund in compliance again with the broker's risk model.

Slight improvement. Although I used Q500US because Q does not scale with QTradableStocksUS.

@Alan I'm choosing the other way: Find new high quality data instead of processing lots of noise.

Hi @Albert just out of curiosity, are you using a free dataset?

Thanks in advance

@Marc
I am collecting/scraping the data myself, it is not available on any platform or as an API.

One of my contest models has hovered in the 10th to 30th place range for last 6 months or so, until Corona and then it shot up to 3rd place. now its dropped down to the 60's. looking at backtest orders my quick takeaway is, the shorts payed out with skewed positive returns during panic. Also comparing the top scores being in the 2's on avg, when before top was around 1.6, I'd assume that those models shorts payed off as well. So I'de guess, your model didn't have good shorts vs your longs.