Regressing Stock Volume on News Volume During Trading Hours
Summary
For a regression relating stock trading volume to news volume, the document recommends using observations from market-open periods and excluding closed-market intervals. Adding closed hours as zero-volume observations can create a misleading relationship: both variables may be zero because the market is closed, rather than because news volume explains trading volume.
The proposed model treats the relationship between news and stock volume as meaningful only while the exchange is open. It also presents an indicator-based formulation that distinguishes open and closed periods, noting that with a linear function and intercept it yields the same parameter estimates as fitting only open-market observations. The answer does not provide empirical results or discuss other modeling choices, such as intraday seasonality, serial dependence, or how news timestamps should be aligned with trading intervals.
Key ideas
- Exclude market-closed intervals from a regression intended to estimate the relationship during trading hours.
- Zero observations for both volume measures can conflate market closure with a genuine news-volume relationship.
- An indicator for open versus closed periods can express when the relationship is assumed to apply.
- With a linear model and intercept, the stated indicator formulation matches estimates from trading-hour observations alone.
Tags
Full text
# Regression of TAQ half-hourly stock volume data against news volume
# Regression of TAQ half-hourly stock volume data against news volume
I am planning to run regression of half-hourly stock volume against the half-hourly news volume for that particular stock. I am looking at 2 years of data for my analysis. However, I am stuck thinking about what should be done to the non-trading hours period on each day?
To be specific: 1. Should I regress the data only for the working hour of the exchange, which means that the Y -values in my regression will contain the "stock volume" in each 30 minutes from 9:30-16:00 on each day from start date to the end date of my regression period and X-values will be the corresponding "news volume" in each 30 minutes?
OR
- Do I need to make the data evenly spaced in 30 minutes and include the "non-trading hours" for each day with "zeros" as the stock volume and the news volume?
I believe the regression result will be different in both the cases. Need an urgent advise.
## Answer by user2763361 (score 2, accepted)
https://quant.stackexchange.com/a/11507
Do not run the zeros against the zeros. This is similar to how weekends are treated in academic studies. There is not five days with two additional days of 0 in the regressions for each week in the sample... there is just the five days (although I do encourage you to read about the weekend effect).
Your hypothesis is that there exists a function $Volume(t) = f(News(t)) + e(t)$. When the market is closed, no such function can exist, so what are you supposedly estimating with the zeros in the regression equation? If you include the zeroes, then what you are saying to the model is that during these times $Volume(t)=0$ because $News(t)=0$. Yet we know this is false, and that they are both zero because $t \in \{Market Close\}$.
If you are really concerned about the irregularly spaced time series, you could consider a more legitimate data generating process:
$$ Volume(t) = f(News(t))*I(t \in \{Market Open\}) + c*I(t \in \{Market Close\}) + e(t)$$
where $I$ is an indicator function. However you will notice that this will give you identical parameter estimates (if $f$ is linear with an intercept) as if you simply estimated the original equation during trading hours only.Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.