Back to Community
How can I trust 51% annualized alpha (!!) from AlphaLens?

I tried a very simple strategy to predict factors and relative displacements of Energy sector stocks w.r.t factors. The results are too good to be true. The alpha does not decay. Obviously, there is something wrong. Could anyone help me, please? Here is the link

10 responses

Alpha does appear to be decaying, no?

FYI, after the notebook (NB) has completed, if you delete the cell(s) that contains your factor calculation and then save the NB, you can share your Alphalens NB without revealing your alpha factor. Might be easier for people to comment on a more complete picture.

Thanks, Joakim,

Please see the attached notebook. When I compute P/L manually the P/L is grossly inverted. Whereas AlphaLens reports significant alpha.

Loading notebook preview...

Your P&L algo may cover a different period? Sounds like you may need to invert the signal otherwise? But of course volatility drag takes its toll.

Your strategy has has a low autocorrelation. Thus, your algorithm has high turnover. Trading cost will eat away at most of your profits.

I figured out the issue. Smoothened alpha to reduce turnover and backtest looks good.

Hi Satya,

Sorry I'm not able to help you on this. I'm just a mere user of AL as well and far from an expert. Maybe have a look at the Alphalens documentation on Github to try to figure it out. Or perhaps a true AL expert (such as @Luca, @Michael Matthews, @Thomas W.) could advise on this?

To me it looks like quite a decent factor (including autocorrelation and turnover), at least as a starting point (appears to be a shorter term mean-reversion type factor). Perhaps the returns discrepancies are due to extreme outliers? Maybe try to 'winsorize' the factor by x% on each side? Maybe also plot the factor distribution to see if there are 'fat tails' or extreme outliers?

See this post and help from Michael Matthews in a similar issue I was having. Unfortunately Delaney has moved on since then, so the thread is somewhat 'dead' unfortunately, but maybe you can find some solution hints to your issue there?

@Joakim. Thanks for the reference. I will review the factors as suggested.

One round trip costs you 28% a year at 5 bps trading cost per side. In the real world costs can more than double, if you need to trade around the open. Even if your strategy didn't get overcome trading cost, you can still use your factor to give your longer term factors a little boost in Sharpe ratio.

Thanks, Indigo. I guess I will have to collect many long term and short term signals before combining them.

if the loss does not come from the commission, your alpha reverted can be a good signal :)