Direction of Change Forecasting II: The case of the UK

[This article was first published on unstarched» R, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

In the previous blog article I discussed a dynamic binary model for the directional forecast of the US equity market using a select number of economic, fundamental and technical variables as predictors. A natural direction for extending that research would be to look at similar models in different countries or for other asset classes. In this article I examine the case of the UK and find some interesting and perhaps surprising results. I also provide a short summary of my frustrating attempts at obtaining quality data as a warning to others who may wish to embark on a similar adventure.

The Data…or lack of it

There is a lot to be said about a society that embraces and promotes open and unfettered access to quality data. When I was collecting data for the US model, it filled me with confidence and satisfaction to find easily accessible long history vintage series of the CPI and Industrial Production via the ALFRED website. Even fundamental index data was available openly from the S&P website which was easy to combine with the long history maintained by Bob Shiller. The same cannot be said of my experience with UK data. The Office of National Statistics ( ONS ), was a complete nightmare to search, contained little in terms of historical depth and forget about vintage point in time series (at least none that I could find within a reasonable amount of time searching). Interest rate data from the Bank of England also lacked the kind of historical depth one finds in the equivalent US series. Instead, I had to rely on the OECD website for both economic and rates data, neither of which were in vintage form whilst the latter was only available with historical depth as estimates (until Dec 1987). The worst experience however was tracking down a continuous UK equity index with a long history. Not only do the FTSE indices not openly provide such data, obtaining dividend and earnings for their indices is almost impossible without a pricey subscription to some institutional service. While their site does provide the last 5 years of data, this must be accessed day by day as an individual pdf format making collection of such tedious and prohibitive (update: the situation it appears is not limited to the UK but also to South Africa where the FTSE has also paywalled the national stock market index, the FTSE/JSE…what a dismal state of affairs).

After much searching, I was able to stitch together a ‘long’ historical series for the FT-All Share index which was a patchwork of the FT 500[1] index from April-1962 to Oct-1972 and the FTSE-All Share from Nov-1972-present. The same was done for the dividend yield. The ONS used to publish the Dividend Yield for the All Share index up until 2011, at which time they stopped (and also removed the data). I am guessing the FT wanted to close access to even this information (what a surprise!). In any case, I had access to that series and manually brought it up to date.

For reference, the dataset is provided here.

Noisy Predictors and Noisy Response Variable

Having collected/patched together a long monthly series for the response variable (FT-All Share Index), economic data on inflation (CPI) and industrial production (IP), rates data for constructing the yield spread, and fundamental data based on the Dividend Yield I embarked on the backtest of the model, as in the case of the US.

The results obtained were not much better than a Buy and Hold strategy. In fact, it would appear that the predictors I used had little predictive ability to capture changes in the direction of the index, at least for the earlier part of the period tested. This was not too surprising. I did not have a lot of confidence in the data, particularly the early part of the dataset which consisted of estimates and values spliced together from different sources. In addition, because I did not have the point in time release data for some of the variables I may have been introducing too large a lag in my attempt to avoid any look-ahead bias. Finally, it could also be the case that the actual market direction, as captured by the FT-All Share index which I had collected and patched together, was also measured with a lot of noise. I do know that some of the previous studies on the UK did find predictive value in the dataset I was trying to use, such as Guidolin et al (2009)[2] , but I do wonder how carefully those studies were are able to align the data with the response variable dates so that there was no look ahead bias, given the historical absence of point in time estimates. For instance, the CPI for a particular month is not released until at least the middle of the next month, while industrial production has an even longer lag.

Timing the UK Market with US Signals

Having found no way around the data problem and the cause of the bad results from the backtest, I decided to see whether the US model signals could be used instead to time the UK market. The results are shown in Table 1 and Figures 1 and 2 below, where I have also included the FTSE-100 and MSCI UK iShares index when they began trading. The results are partly surprising since I did not expect that the US signals could provide such a good timing tool for the UK market[3] . The natural next step would then be to test the UK market direction using the US predictors. Unfortunately, this did not yield the same results as using just the US signals which pointed to a problem with the response variable. Smoothing the FT-All Share Price using a Zero-Lag EMA prior to generating a directional response variable did improve the results somewhat, providing some validation of the noise hypothesis, but again proved inferior to just using the US signal.

1979-2013
DBM (FT-ALL)
EMA (FT-ALL)
B&H (FT-ALL)
 
1996-2013
DBM (EWU)
DBM (FT-ALL)
DBM (FTSE100)
EMA (FTSE100)
B&H (FTSE100)
CAGR12.357.638.3CAGR11.359.49.595.253.14
Vol(Ann)10.5212.3816.04Vol(Ann)14.139.6811.179.7815.57
%Up79.7575.1963.21%Up77.8877.477.474.0459.62
MaxDraw16.1934.4846.87MaxDraw22.8515.8716.5617.4346.64
CAPM(alpha)0.05490.01CAPM(alpha)0.06610.06040.06290.0232
CAPM(beta)0.43580.6064CAPM(beta)0.57160.46480.53130.3955
Timing3.75120.6409Timing2.70945.55863.69550.8057
Sharpe0.67810.22180.2026Sharpe0.60080.68020.60680.25980.031
Information0.71090.2320.2124Information0.61470.69490.62040.26650.032
Calmar0.76290.22140.177Calmar0.49650.59210.57930.30120.0674
Kurt(ex)1.71899.60693.1604Kurt(ex)4.61381.62813.74972.19270.6434
Skew0.0257-1.5463-0.8452Skew0.5224-0.10220.1791-0.5664-0.4183
1984-2013
DBM (FT-ALL)
DBM (FTSE100)
EMA (FTSE100)
B&H (FTSE100)
 
2010-2013
DBM (EWU)
DBM (FT-ALL)
DBM (FTSE100)
EMA (FTSE100)
B&H (FTSE100)
CAGR10.219.795.396.13CAGR12.579.8810.921.54.69
Vol(Ann)10.5511.2111.916.42Vol(Ann)16.1111.3811.5910.6914.9
%Up77.8477.8473.160.8%Up77.2770.4572.7365.9159.09
MaxDraw16.1916.563346.64MaxDraw22.8515.8716.5612.3916.56
CAPM(alpha)0.04740.04460.0044CAPM(alpha)0.07880.05350.0747-0.0088
CAPM(beta)0.44710.4740.5243CAPM(beta)0.62090.69620.62790.5122
Timing3.34232.87270.3768Timing3.60132.2045.37690.4235
Sharpe0.5690.49930.12620.1257Sharpe0.77630.8630.93680.13480.3107
Information0.58890.51690.13120.1308Information0.77690.86380.93770.13490.3109
Calmar0.63060.59110.16320.1314Calmar0.55020.62270.65940.12090.283
Kurt(ex)1.79272.987412.54683.3172Kurt(ex)3.00310.52562.19051.40980.353
Skew0.00550.1957-1.7455-0.6945Skew0.76810.19080.7789-0.4230.1033

 

Figure 1

Figure 1

 

Figure 2

Figure 2

Comparing the US and UK Predictors

As a final part to this story, I’ve provided comparative plots of like for like variables for the US and UK markets in Figures 3 to 5. The UK 3 Month T-Bill does appear to be much more noisy than the US one, and it is interesting to note how the 2 spreads converge and track so closely post 2009 (though the individual rates are close to their lower policy bounds). The UK CPI and IP also appear to contain more ‘noise’ especially for the early part of the series, which as mentioned was based on OECD estimates.

Figure 3

Figure 3

 

Figure 4

Figure 4

 

Figure 5

Figure 5

 

Concluding Remarks

I am still uncertain as to how to interpret the results. Perhaps with a cleaner dataset I may be able to make more concrete inference about the UK case. For now, the US signals provide an adequate timing mechanism for the UK equity market.

If you have some unique insights into any of the questions raised, feel free to drop me an email.

[1] The FTSE 500 was discontinued in 1992.

[2] The researchers had access to datastream and global financial database, the latter quoting an exorbitant price for just the total return FT-All share index which I refuse to pay.

[3] In order to avoid any look ahead bias because of the later closing time of the US market, the results for the FTSE-100 and EWU are based on the returns the day after the signal was produced in the US model (for the FT-ALL Share I only had monthly data so could not capture this).

References

Guidolin, M., Hyde, S., McMillan, D., & Ono, S. (2009). Non-linear predictability in stock and bond returns: When and where is it exploitable?. International Journal of Forecasting, 25(2), 373-399

To leave a comment for the author, please follow the link and comment on their blog: unstarched» R.

R-bloggers.com offers daily e-mail updates about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

Never miss an update!
Subscribe to R-bloggers to receive
e-mails with the latest R posts.
(You will not see this message again.)

Click here to close (This popup will not appear again)