October 1, 2008 8:50 AM | Posted by mhf: | Reply

Why are you so stuck on CNBC? Are you on the payroll at Fox? Is CNBC reporting different information than the rest of the media?

I think the Bush administration has done its best to panic people about the economy in the weeks before the election; they used a similar tactic/strategy (I don't know the difference, either) with terrorism before the 2004 election.

It's difficult for the general public to interpret all the economic news we're being inundated with (by all media sources); the economy really is in trouble, and many of us are afraid of the impact on our personal lives. I think the market tends to reflect Wall Street's interpretation of events as reported by a variety of sources, and as viewed through individuals' particular political lenses.

You seem to be focusing rather heavily on one media source to the exclusion of all others. What if you used some different outlets as control groups? This seems very unscientific.

Score: 0 (4 votes cast)

October 1, 2008 9:14 AM | Posted by mhf: | Reply

For a different take on the stock market, please see Turboglacier's blog, May Shrink or Fade

Score: 1 (1 votes cast)

October 1, 2008 8:37 PM | Posted by AK: | Reply

Hmmm, what about the old principle that, by itself, correalation doesnt prove direct causality?

We gotta come up with a research design that would enable us to test whether there is a direct cause and effect between stock market performance and CNBC ratings or whether there's a third variable that could be affecting matters.

Beyond that, I dont know what test to use, maybe a well designed multiple regression model?

Score: 0 (2 votes cast)

October 3, 2008 1:09 PM | Posted by MedsVsTherapy: | Reply

A simple correlation is fine. My guess is you are correlating each day's CNBC rating with each day's volatility index? If you suspect a predictor pattern, of two days, then you could also correlate (Pearson r) the CNBC rating of one day, say Wed Oct 01, with the VIX of two days later, say Fri Oct 03. The resulting "r" would be greater (although you can't get much greater than .889).

An alternate strategy: regress VIX onto CNBC: like this: the old-school 'formula for a line:'
VIX = (some coefficient) * CNBC.
"coefficient" means the same thing as "weight:" if you prefer a restaurant based 3/5 on ambience and 2/5 on price, then you would get the ambience rating, multiply it by 3/5, and get the price rating, and multiply it by 2/5, and you will have your overall preference for that specific restaurant.

You ask your stats program to give you the unstandardized as well as the standardized coefficient. The standardized coefficient is "interpreted" exactly like the Pearson r: how close to a perfect, one-to-one correlation is there? Like Pearson r, it will range from negative 1 to positive 1. Absolute value closer to 1, away from 0, is stronger.

With the unstandardized weight, or unstandardized coefficient, you multiply a given day's CNBC score by the coefficient and you get your predicted VIX. That would be the benefit of performing a few more clicks in the stats program.

Then, hypothetically, if you are a gambling man, you take your CNBC-based prediction for where VIX will be in two days, and place your money accordingly. I guess you get the Wednesday rating some time late Wed or Thursday, and the volatility of the entire Friday is represented by the end-of-day summary volatility index, so really, you only have one day for predicting, from close-of-market Wednesday to when you need to place your money to benefit from some prediction of Friday's volatility, which I guess would be by opening bell Friday morn. So you don't quite get a full 2 days.

Alternately, you place your bet as soon as the CNBC rating comes out, since others will be making siilar moves based on similar info.

With the multiple regression, you also will get whatever confidence interval you want; "default" is usually 95% but you often can select other ranges, i.e., 90% CI, or you can specify your own confidence interval. Of course, to be more confident in a prediction, you have to broaden the range of predicted VIX. similar to: I am 99% sure tomorrow's temp will be between 55 degrees F and 95 degrees F, but only 65% sure it will be between 70 degrees and 85 degrees. So: should I pack a sweater? I have a small but possible risk of the temp being in sweater range, so if I really need to be sure, I take the sweater, but I can risk it so I won't. Same with predicting VIX: you want the most narrow prediction range, but the more narrow, the less "confident" or likely you are to be right, or really close to right, day after day.

You can do a "multiple regression" with only one predictor and one outcome variable. In that case, you just properly call it "linear regression." People rarely do analyses with just one predictor, but it fits in this case, unless you want to add other predictors: price of oil, unemployment rate, etc.

Causality: if one preceeds the other, and the relation holds for a long time - like a year, not just a few days, then I would feel fairly confident that there is some causal connection between these two, and that it is something about CNBC that causally is related to subsequent (two days later) level of VIX. It is difficult to defy the time-space continuum. So, correlation does not infer causation. Neither does correlation with a necessary temporal sequence, but the temporal sequence is a lot better indicator of a potentially causal relation than simultaneous correlation.

Score: 1 (1 votes cast)

October 4, 2008 3:25 AM | Posted by Anonymous: | Reply

You can't use a correlation of the data points themselves because they are not independent, i.e. today's measure is correlated with yesterday's measure. If you doubt this, correlate VIX(t) with VIX(t-1). Perhaps correlate the percent change from one day to the next? It will not eliminate the independence problem completely. There are sophisticated time series analyses for this sort of problem (ARIMA models, I think) but I am not that familiar with them.

Correlation does not imply causation even if there is a time lag. There could still be some underlying causal factor that drives both but at different lag intervals. Such as news. Cherry-picking your data by ignoring the data points during Olympics coverage is cheating. Get yourself a longer time series instead.

Score: 1 (1 votes cast)