Excuse the pun with the post title (Unintentional honestly!)
I have to admit I'm absolutely stumped by the things I'm investigating at the moment.
I would normally give up and move on, but I believe these things have implications for robust adaptive system development.
It concerns the Agg M indicator which I've blogged about a few times.
Aggregate M Indicator
As a quick recap, the Agg M works on the principle that markets are mean reverting in the short term but trend following in the longer term. The Agg M blends these two principles, with the longer term trend helping to smooth the more profitable but volatile short term mean reversion.
I wanted to have a look at making the Agg M adaptive, mainly because it broke down in May.
This is based on the bog standard settings of opening trades at the next close and holding for 1 day. Buy above 0.5 and sell below 0.5. I tested the FTSE 100 and the SPX, results were roughly similar.
'Broke down' is a bit excessive, but especially on the SPX, there were indications in May that the Mean Reversion part of the system was not firing. May saw a huge upsurge in volatility, a market condition that is 'supposed' to good for Mean Reversion, yet the Agg M had its worst month for a long time.
An Adaptive Agg M
I got thinking, well if the Mean Reversion part is falling down, that could be easy to manage and isolate. The indicator is built on the premise that markets are mean reverting in the short term, but what if they stop becoming mean reverting and do something else instead (or the type of mean reversion changes etc?).
So how do we track if the system is starting to become unprofitable?
My intimidate thought was to try the Expected Value levels on a rolling basis to see if it provides an means of switching the entire system on/ off or at least the mean reversion part.
To cut a long (and very boring) story short, I tried many different combination and rolling EV lengths, but applying the filter based on system performance tended to decrease overall performance, not increase it.
The chart below shows the problem visually. The problem with the Agg M, even with the recent under performance is that it is never consistently unprofitable. If anything, there is a rhythm whereby it bounces from weak performance to strong performance and back again. The data is of the FTSE 100 from 1997 to 2010.
A score of 0.5 or above means the system has been profitable for the last 50 periods.
I ran a rolling 50 period correlation comparing today's Expected Value Score with the Expected Value score from 50 periods ago.
The correlation strength waxes and wanes, but the clear bias is for the correlation to be negative. In other words a strong performance 50 periods ago is more likely to mean a weak performance today. Conversely, a weak performance 50 periods ago is more likely to mean a strong performance today.
This makes using an Expected Value filter very hard because each time the system looked like it was about to break down, it was actually more likely to go on and make bigger profits.
Correlation does not equal causation, but there appears to have been a relationship there.
Can we use this?
I ran a filter as follows:
Trade the indicator as normal except when today's rolling Expected Value is 50% greater than 50 periods ago. In addition, today's Expected Value also has to be greater than yesterdays. Logically with a negative correlation, this should improve performance because when the EV is 50% higher than the past, the more likely outcome is for the system to under perform, so we cut these trades out.
AggMA = The Adaptive Agg M.
Slightly better performance in 2010, but a better total performance.
While interesting and potentially useful, the findings to not answer my original question on how to tell when Mean Reversion is falling apart. Neither does it provide a good way to finally say that a system has bit the dust. It seems a good way of working out when a method has over performed in the short term, but not as a switch on or switch off, which is what I was looking for.
- Indicators/ methods, just like markets have their own rhythms. Perhaps a completely unprofitable method could be made profitable by exploring its rhythm.
- Perhaps measuring the number of consecutive days a method goes with an EV under 0.5 might help as an off switch.
- Building adaptive systems is bloody difficult!