Market intraday momentum

I recently spotted the following intriguing paper: Market intraday momentum.
From the abstract of that paper:

Based on high frequency S&P 500 exchange-traded fund (ETF) data from 1993–2013, we show an intraday momentum pattern: the first half-hour return on the market as measured from the previous day’s market close predicts the last half-hour return. This predictability, which is both statistically and economically significant is stronger on more volatile days, on higher volume days, on recession days, and on major macroeconomic news release days.

Nice! Looks like we can all become rich now. I mean, given how it’s written, it should be quite easy for any individual with a trading account and a mouse to leverage up and start accumulating. Maybe this is so, but let’s have an informal closer look, with as little effort as possible, and see if there is anything we can say about this idea.

Long ago I used to day-trade, at some point I collected intraday data for my own backtesting and placed it online for replication of a post about realized volatility. The data is an array with dimension [390, 8, 250]. The 250 refers to days: 250 trading days ending in 22/10/2012. The 8 refers to the number of columns which are basically the open, low, high close, volume etc. The 390 refers to the number of minutes in each day. We can load it in R using the load(file = ".../prlev2.RData")*. The ticker is the SPY ETF. It’s only one year but let’s see if the price behavior of that year supports the claim in the aforementioned Market intraday momentum paper.

It looks like that:
Market intraday momentum - half and hour returns To examine whether there is a predictive value we can plot the last 30 mins returns over the first 30 mins (remember those include also the overnight returns because they are calculated based on the close of the previous day, hence the wider dispersion in returns).
last 30 mins over first 30 mins The figure supports the claim, and a simple regression model returns:

So significant slope indicates that if the first half an hour of the returns are up by one percent, and if all the usual assumptions hold (linearity etc.) then on average we should see the returns of the last 30 minutes of the day up by 0.06%, which may give sufficient edge for money-making. Especially in this specific heavily-traded SPY ticker.

The However we have all been waiting for

In statistics there is the concept of “influence”. The math is quite ugly (wikipedia link) but the concept is super simple and super important. Which data-points drive the results? The way to measure influence is by using something of an “in-contrast” principle: what happen to the results if we remove this or that observation? If the results don’t change much, the observation is then told off for being uninfluential*. This “influence” concept is somewhat related to resistant regression. Robust regression tries to dampen the influence of outliers. Resistant regression estimates the parameters by altogether removing any influence from outliers. There are few methods we can use to perform resistant regression. We use here the least quantile of squares method as we did last time. It is simple enough to implement. Formally:

    \[\widehat{\beta} = \underset{\beta}{argmin \;} \varepsilon_{ (\nu)}^2 ,\]

so we find the estimate which minimizes the squared order residual, but only those residuals which belong to the \nu region. The \nu region is selected as the arguably most representative of where the data is expected to lie. The region \nu in our case would exclude those observations which are at the top and bottom k percent, with k being 3% and 5% for example. This is plot of the new resulting regression lines from using the least quantile of squares method:

least quantile of squares regression
Here is the function needed to compute those regression lines:

As you can see, when k= 0.03 you lose significance, and when k= 0.05 the line is flat. Clearly the results are driven by a small percentage of the data. Looking again at the abstract, the authors acknowledge that finding, in a way: “… stronger on more volatile days, on higher volume days, on recession days, and on major macroeconomic news release days.”

Conclusion

So if the results may be driven by such a small percentage of data points, I doubt that a full blown backtesting exercise would yield any meaningful tradeable strategy. There can be very many months of random-walking at best, which can be costly. That said, momentum is known to work interday so it can be that it is extendable to intraday, especially when the returns of the first 30 mins include the overnight returns. The Journal of Financial Economics is a top journal in the field, and so there is no doubt that the paper contains meaningful contribution. But the contribution is probably not “do this and get rich”. What is shown above is a flat, minimum effort statistical analysis and should be taken only as such.

Footnotes

* UPDATE: some people had issues with the data. You can download it to your computer from here and load it to R from the relevant directory.
* It is a bit more complicated since “results” are not yet well-defined. We usually speak about changes in particular parameters, an observation can have different influence on different parameters.

One comment on “Market intraday momentum”

Leave a Reply

Your email address will not be published.