“Sell in May” Debunked?
This time every year the “sell in May and go away” strategy rears its head again.
It’s hard to buy in to the idea that such a simple approach could have such divining powers, but the results are (on the surface) compelling.
I’ve been pondering how best to put the strategy through the paces. Thoughts…
[growth of $1, logarithmically-scaled]
Usually a graph like the one above accompanies these discussions. Here I’ve shown the S&P 500 (dividend-adjusted) from Nov-Apr (red) vs May-Oct (grey), since 1950.
Awesome results. Great strategy.
The problem is of course that this is all prepared with the benefit of hindsight. Surely in 1950, we wouldn’t have known that Nov-Apr would turn out to be such fortuitous months for stocks. So in the next graph I’ve taken a different approach.
I’ve assumed that each year the investor only looked at the data available from 1930 up to that point in time, and invested in whatever 6 months of the year had been the best for stocks.
This is called “walking the test forward”, and (to some degree) removes the benefit of hindsight.
[growth of $1, logarithmically-scaled]
The graph shows that most of the benefit of choosing seasonally strong months disappears because the investor wouldn’t have made the “right” choices given the information available at that time.
The investor would have done well since the 1990’s, but that’s a much less robust observation than the first graph would imply.
. . . . .
So what if rather than choosing seasonally strong months based on ALL data available up to that point in time, the investor only looked at say the last 10 years?
Same conclusion.
20 years?
Same conclusion.
If we go out to about 30 years (i.e. the investor is choosing seasonally strong months based on the previous 30 years of S&P 500 data), the strategy soars again…
But the fact that only 30 years (as opposed to say, 20) works so well is most likely because it’s a curve-fit solution.
So does the data totally debunk “sell in May”?
No. I wouldn’t base a trading decision solely on the rule, but results in all tests were impressive enough in recent history that the observation at least deserves to be on the radar.
But that really misses what I think is the more important point:
The graph like the first I showed would lead the reader to think that the “sell in May” rule is much more robust than it actually is. In truth the rule is at best a questionable observation, and at worst, simply a product of randomness.
Happy Trading,
ms
P.S. This post isn’t meant to dump on the quantitative minds who I respect very much that have discussed this subject recently. I’m just one nerd with $0.02 and I recognize that on this one, I am probably out on a long branch all by myself.
. . . . .
To stay up to date with what’s happening at the MarketSci Blog, we recommend subscribing to our RSS Feed or Email Feed.
Filed under: Time-based, Trading Strategies | 15 Comments
I would like to see full year vs Nov-Apr historically. Would also be curious to see how it looks if you tweak the start date (Oct). Can this data set be made available for us to play with?
Hello Chuck – you can see full year versus Nov-Apr + investing in bonds during the off months here:
http://marketsci.wordpress.com/2012/05/02/econompics-seasonality-strategy/
Any other 6-month combo is going to be somewhere south of those results.
I hadn’t planned on showing any other alternative 6-month periods, but if you’re interested, S&P 500 data is readily available online (but usually needs to be dividend-adjusted).
michael
Posts like this that make me especially happy you’re blogging regularly again. Thanks for the insight!
Great insight. The most important takeaway (in my opinion) is your point on “walking the test forward.” Analyzing data this way is far superior to the data mining practice of optimizing over the full data set…just my two cents.
Promptly de-bunked! I absolutely hate when CNBC slavishes their attention on this strategy.
Hello John – I wouldn’t say totally debunked – the fact that the walk forward is a dud doesn’t mean the observation isn’t “real” today, it just means that trading the best performing 6 months hasn’t worked so well historically.
Given the consistency of Nov-Apr over the last couple of decades, I leave open the possibility it’s the real deal.
I think the more important takeaway is that showing performance of “sell in May” for the last 40, 50, or 60 years (like most people do) is debunked b/c it’s clearly hindsight bias.
michael
great post! keep up the good work
Thanks so much for the post – was vaguely worried after the first post on Selling in May strategy and wanted to do a walk forward analysis – but you did that for us – such a relief !!!
Thanks again
Thanks for your interesting post.
A good out of sample test would be to see who first published the idea, then test “SELL IN MAY” from that point forward. Yale Hirsch of “stock trader’s almanac” claims to be first (old edition – http://www.amazon.com/gp/product/0816010447/ref=ox_sc_act_title_1?ie=UTF8&m=A2QOAC51JE63TW), but I haven’t verified this and don’t know the supposed date). I would order this old edition to check it but I live in Canada and they won’t deliver it to me here.
Other work seems to show it works in the UK for 300 years (Are Monthly Seasonals Real? A Three Century Perspective) and for something like 35 years in 36 out of 37 countries examined
(Sven Bouman; Ben Jacobsen (2002). “The Halloween Indicator, “Sell in May and Go Away”: Another Puzzle”. American Economic Review 92 (5): 1618. ). of course country returns aren’t uncorrelated, but they’re far from perfectly correlated (especially earlier in this history).
Hello Matt K – smart comment – I was waiting for someone to point out that “sell in May” was first suggested long ago (I too don’t know when it was first suggested).
IMHO, this is a more subtle version of hindsight bias.
People make observations and predictions all the time. Just because one happens to pan out in the future doesn’t necessarily indicate that the prediction was sound (only that randomness shone favorably).
In my mind, if you can’t show with a walk-forward that someone would have reasonably come to that conclusion at that moment in history using the same logic that’s being used to justify the conclusion today, then the conclusion is bunk.
I also think the fact that the observation exists in most equity markets is a red herring. Most equity markets (even prior to the computer age and what we think of today as globalization) were correlated, so it’s no surprise that what’s good for the goose is good for the gander.
Again, smart comment and I appreciate the thoughts.
michael
P.S. your comment inspired me to take a extend my test back to 1871 using Shiller’s dataset (http://www.econ.yale.edu/~shiller/data.htm).
I’m not showing the results on the blog b/c I think I’ve beaten this horse to death, but I do think my conclusion holds (and if anything, this extended dataset makes the conclusion even stronger).
michael
Thanks Michael for your comments and added testing!
Let me start by saying I’m not a ‘true believer’ on this topic, but do feel this one is still somewhat more of a ‘shade of grey’ than you’re depicting it. So, briefly:
On the CON side for this idea:
* a basic ingredient for any good anomaly is a good story behind it, and this one is weakest here – some published nibbles (though I won’t claim to have dug in exhaustively).
* trends in short term interest rates might explain when this idea worked, for the times it did work, so seasonality could be spurious
On the PRO side (PROs and CONs aren’t unbalanced because I’m PRO):
* Correlations across equity markets in earlier decades (e.g., 70s, 80s) were much lower than today, so cross-market robustness is relevant.
* Going back to the original source of the idea, trying to actually find out their motivation (vs speculating on it) and testing from that point forward isn’t hindsight bias, it’s good research practice.
* If enough people believe in the idea it might tend to work better over time for that reason (think Andrew Lo adaptive market hypothesis; try a google trends analysis on “sell in may”) – all the press the idea gets could itself be contributing to the effect.
* Your typical anomaly doesn’t work as well if you test it much further back because there is on average some amount of data mining bias in published work (but we can’t test most that far back due to data limits). So that’s a higher bar being applied here vs other ideas.
* I’ve seen published work before on this back to ~1870. One thing to keep in mind is that that the US was in recessions/depressions for much/most of the1870-1900 period, and many ideas perform differently in calm vs manic markets (e.g., value vs momentum in stock selection); 1914-18 and 30-45 might be unusual for similar reasons (and see the 300 yr UK study reference above).
* I recall May-October as having lower volatility than the rest of the year, volatility has smaller error bars and is less subject to data mining as a result, so this idea might look better on a Sharpe Ratio basis than a simple total return basis.
OK, I’ll stop now.
Hello Matt – thanks for the thoughts, especially re: the potential for self-fulfilling prophecy. I completely agree.
One last thought – I do think you hold ideas like this to a much higher standard than ones that occur more frequently (daily, weekly, etc) because we’re really talking about a very small number of observations (my test all the way back to 1950 in a sense includes a mere 60 observations).
So I do think more extensive tests like the walk-forward are justified and I’m unimpressed that this particular prediction (out of the infinite number that are tossed out by pundits such as myself) has performed so well in the last couple of decades (because again, that encompasses so few observations).
It’s like that old market adage about the pundit who makes a 1000 random calls. Half are right and to those 500 he makes another call. Half of those are right and to those 250 he makes another call. This goes on until 1 guy thinks the pundit is a genius.
It takes a lot to convince me that these types infrequent observations are anything but that type of “fooled by randomness” scenario.
Just my $0.02. Thanks again for the thoughts (and for blowing up my comments section =)
michael
Hi Michael,
Thanks again on your comments, and welcome for blowing up the comments section :-). I very much agree you want something that controls for data mining risk (1000 random calls, then 500, etc.). Testing starting from the date of first publication is one way to do that, as does having appropriate statistical tests to apply to the results. This particular case is more complicated than some!
matt
It’s not dumping on quantitative analysts when you share findings based on equally discerning quantitative research. Nice work…thanks for sharing.