I am working with intraday stock prices. I have found that the autocorrelation between the returns is negative (significantly so, but the value is very small). I am aware of how to interpret negative correlations from a technical point of view, but what could be the real-life reasons for that?
1
-
$\begingroup$ As I commented under the accepted answer, it's not clear how universal/significant this effect is and what data frequency you use. I would imagine any kind of stable autocorrelation effect would've been exploited already. Tick by tick market orders series often shows positive autocorrelation as opposed to negative. any other level of data is somewhat aggregated and the effect of autocorrelation is suspicious. If I were you, I'd do the same analysis for a bunch of stocks and a set of days. If the stats on all these samples are significant, then there could really be something to discuss about. $\endgroup$– Will GuCommented Dec 9, 2016 at 18:40
Add a comment
|
4 Answers
$\begingroup$
$\endgroup$
1
Looking at transaction prices, they would occur at the market bid if the active part is a seller, and at the ask if the active part is a buyer. With a random flow of sellers and buyers, the price will bounce between the bid and ask prices, creating a negative autocorrelation in returns.
This penomenon is known as the bid-ask bounce, and has been discussed here in the follow up to this question on how to adjust for the bounce.
-
1$\begingroup$ At tick level, this is not (necessarily) true. Market orders cluster -- it's often observed that buys are followed by buys and sells are followed by sells, possibly a result of execution. It's unclear what frequency, time range and stock name OP was looking at, nor was it clear what "significantly so, but the value is very small" means. For mid frequency data, this could be some kind of mean reversion. $\endgroup$– Will GuCommented Dec 9, 2016 at 17:56
$\begingroup$
$\endgroup$
Check the paper by Cont (2001). There is a stylized fact about it. Autocorrelation should be absent unless for very small time periods, such as intraday (5 in 5 minutes).
I don't think there's a really good explanation, since intraday stock prices follow a random walk, so negative autocorrelation is possible. Perhaps one could say that the performance of the company (financials) vs. the news about the company (speculation) can cause negative autocorrelation.
$\begingroup$
$\endgroup$
1
One possible reason for negative autocorrelation is an imbalance between dynamic option hedgers.
Say an oil producer wants to buy puts on oil, a bank will sell it to them and then delta hedge the risk, whereas the producer won't.
So, if the price of oil goes up the delta of the option for the bank will decrease (as the option goes out of the money) and so they'll sell oil into the market until they hit delta-neutrality, putting selling pressure while the oil producer won't be hedging so the impact is only coming from one side. And viceversa for a decrease.
Another possible reason is that it is hitting stops both above and below. So the price rises a bit hits a stop, comes back down hits another stop bounces up again, etc. etc. until it clears all the stops on one side and then generally rips through.
-
$\begingroup$ "so they'll sell oil into the market until they hit delta-neutrality", shouldn't it be buy, since the bank has already delta hedged the put by shorting the stock, and while the option goes OTM, they reduce their short position? $\endgroup$ Commented Dec 10, 2020 at 22:45
$\begingroup$
$\endgroup$
Negative autocorrelations are typical in high-frequency datasets. Bid-ask bounce is the classic reason given, but it doesn't add up without some extra structure. Another intuitive answer is provided by Cont and de Larrard (2013, SIAM J. Financial Math.): shortly after an upward price move, the number of shares at the new best bid is often smaller than the number of shares at the new best ask queue, because the new best bid consists of orders since the price increase, while the new best ask consists of orders previously sitting at the second best ask price. If orders come in randomly, then the new best bid queue is more likely to be depleted than the new best ask, which corresponds to a decrease in (mid) price.
