Hi Johnny,

Does it exceed 40% even if you set max_short_term_reversal to close to 0.0? (I'm assuming it's long exposure).

If so, I'm guessing it might be your weekly rebalancing that might be causing this. Have you tried rebalancing daily, and instead subsample your factor on a weekly basis? I've never actually done this in the IDE, so I don't know if it's possible or not (and I'm no expert on it, I barely know what it does), but I would think it would need to be done outside of Pipeline, perhaps in the rebalance() function? Something like this maybe (code again from one of Thomas' NBs)?

dts = factor.index.get_level_values(0).drop_duplicates()  
dts_subsampled = dts.to_series().resample('1W-MON').first().tolist()  
dts_subsampled[:5]  
sub_sampled_factor = factor.loc[dts_subsampled]  

@Thomas, @Rene, @Dan, maybe one of you could comment and correct me if I'm wrong?

Hi Joakim,
thank you for suggestions.
I do indeed set max_short_term_reversal (and min) to very low values, but there is almost no effect.
If I re-balance daily the strategy is not as profitable.
The max_short_term_reversal is a factor computed daily(based on 15 days returns) and I think resampling would not have an effect on how the risk exposure is computed to score the strategy(daily I think).

Somehow the optimization done at the beginning of the week is not well preserved during the week, but I was wondering if there could be some other way to control it. With daily re-balancing it can normally be controlled. I tried to re-balance every 2 day for example and the profitability goes done a lot.

Thanks

Hi Johnny,

Looks like, by construction, your alpha signal give too much weight on the short-term reversal factor. Thats why the optimization fails. Try to either mitigate your alpha signal with another factor or relax some filters

Adding another alpha works, but I would have preferred to avoid it. In general the style exposure is not really a risk but is seems more a away to avoid strategies based on well known factors, such as Fama French. In other words Quantopian wants new strategies, bur probably is already using the well known ones, and does not want any replication. I already talked too much ..

I do not agree. Factors are real risk pockets that every equity fund manager have to monitor. It is exactly like other concentration risks such as geographical, sectorial... If you put too much exposure on a single sector and that sector experience a slowdown, what happen to your portfolio ? it will underperform well diversified ones. The same apply for factors. If you put too much weight on growth factor your portfolio will tend to outperform the market during certain economic and market periods (97-2000, 2009-today), it will underperform during others (2001-2009 for example). Same for volatility biased portfolios, vol biased ptfs will outperform during strong and stable economic & upward sloping markets, but will crash deeper during financial stress periods, etc.
This is even more important here as the objective is to be neutral (market, dollar, beta, sector, factor...), if it is not the case, your alpha is not pure alpha and your algo returns will not be consistent in each market cycles.

Short term mean reversion has been the basis of hedge fund strategies for a long time, but know it just does not work as well because it is of public domain. Same for momentum. Their profitability depends on market regimes as you mentioned, but in themselves they are not risks, but but rather as the names says it, "styles". Being neutral to these styles just means to have a strategy which produces profits not strongly correlated with those other "classical" strategies, but the latter are not necessarily more risky. They are just well known and quantopian would not need to invest so much money to make the platform free, if interested in implementing them. They want some new alpha.

Again, thinking this way is an error imo. Short term mean reversion bias simply means that you take the bet that the trend will revert. This is simply the opposite of trend following, which is the bet that the trend will continue in the future (or momentum). Both are bets, and when you take bets you’re at risk. What Q (and most hedge funds) is searching is market distorsions created by any reason (behavioural biases for exemple) that you can exploit on a risk-free basis, or at least a risk that is less than expected returns that the strategy produces (i.e. Positive risk-adjusted ratios), otherwise it has no value imo.

Some of Quantopian styles ,Size and value, are Fama French factors, which are well known to explain a large component of stocks returns. In other words are good predictors, or in the language of quantopian, good alpha factors (to distinguish them from another Frama French variable, market returns, whose correlation with stocks is taken into account by the beta parameter).
I prefer to just call all these variable predictors, which is general terms used in time series analysis.
Momentum and short term reversal are also well know predictors.
Being neutral to them does not mean to have less risk, it just means to be get returns which are no correlated to those predictors (I think called specific returns by quantopian) . The reason is quite simple: who would start a company to re-invent the wheel?

Let's say you found a new alpha factor. If it becomes of public domain, Quantopian could include it in the list of styles, make money on it without informing you, and force everybody to develop new strategies not relying on it. Of course this is an extreme example, but if you were the one who discovered size and value predictors, it could have happened to you.

From the point of view of Quantopian this is a risk management strategy in the sense that investing in many different strategies which are not correlated can be considered an effective risk management approach.

But no alpha factor in itself is more risky than another., and if you were neutral to all possible factors you would juts make no money .. I guess it can be proved mathematically and rigorously.