When Do Investors Freak Out? Machine Learning Predictions of Panic Selling [pdf]
poseidon01.ssrn.com
poseidon01.ssrn.com
This pisses me off. I don’t want any of my financial firms handing out my data to anyone. It looks like they even gave them demographic information specific enough to likely be identifiable. What firm is this data from? This should be illegal.
I don’t know if I can assuage your concerns but…
- any project like this goes through layers of lawyers on both sides (mit and the brokerage). They are extremely careful about exactly your concern.
- mit has infrastructure to securely hold data. They do a ton of defense work, for example. They take it seriously. In the non-defense context, accidentally exposing certain health data can (iirc) lead to the entire university losing eligibility for NIH funding. At MIT that might be a half billion dollar hit. They don’t mess around with that. (Compare to the private sector where there are effectively no meaningful fines or consequences for data breaches.)
- none of the researchers care about you as an individual enough to try to deidentify you in the data. I work with health data, some of which includes addresses. I have never thought for even one second that I should find out who lives at the address, even when dealing with data which includes the city I live in (so potentially my neighbors, eg.)
- everyone involved in the project also separately promises not to de-identify anyone. Again: I really doubt anyone I have ever met in my field would care identify someone, but we do promise not to.
- any data which is going to be merged with whatever the researchers got from the brokerage will be outlined in great detail in advance.
As another point of comparison, how many breaches of university research data are you aware of? These things happen in the corporate world all the time with extremely sensitive data but I have not heard of university data beaches myself.
Finally, there is generally some scientific benefit to the work that the researchers do. We know something from this paper about panic selling which we didn’t know before. That may be valuable.
Research ethics 101: Don’t use people as lab specimens unless they affirmatively consent. Period.
We should go back to the days of actual ethics in research.
Banks and brokerages can do whatever they want with customer data, subject to relevant laws and contracts. Researchers who ever want a penny from the public should be held to higher ethical standards.
Corps can put whatever they want in tos, and researchers are free to use data from those corps. But federal agencies should refuse to fund those pis and their institutions.
This is (and has always been) waivable in cases (like this one) where either obtaining consent is impractical (which maybe applies here) OR you are making use of de-identified data (more likely to apply in this case). Affirmative consent still is required in many cases.
With health data (the case I know personally), it would not be feasible for me to find five million patients and ask them to participate in my boring study answering a question which wouldn’t interest them. Just mailing each of them a letter to ask would be cost prohibitive!
The value of research is frequently non-zero, whereas the risk to the participants is generally zero or extremely low. This is the balance the IRB is trying to strike.
And believe me, as someone who has gone through it, I think they err way way WAY too far on the side of being overly cautious!
Nearly every large organization has probably had a data breach, the only questions are if it was discovered and if so if it was reported. MIT can have the best cybersecurity team in the world and they will still be helpless against a 0-day.
“None of the researchers care about you” - strawman, no one said that or thinks that, and it’s irrelevant.
“Everyone involved promises not to identify anyone” - see response to first point.
I think the other points are valid though.
They legally promise via some form a non-disclosure agreement. A violation means they are liable in the case they are the source of a breach, which can both provide incentives and recourse. Certainly not ironclad, but it is stronger than just a vague "promise."
First you have to care enough to go after the NDA breaker. Then you have to have enough evidence to legally demonstrate they broke the NDA. Then you have to pursue legal action. This all costs time and money, and even then there's a good chance you can't actually prove anything.
You have to seriously piss off someone (most likely a large corporation) with a lot of time and money for it to matter that you broke an NDA.
OP said he didn’t want to be de-anonymized. Either the researchers do that themselves or the data is stolen and someone else does it. For the first group, I tried to explain why the researchers themselves aren’t going to and why.
So I’m not sure no one thinks it and it’s not irrelevant.
> Either the researchers do that themselves
Has anyone suggested that? I don’t think anyone is concerned about the motives of MIT researchers.
Until you become a public figure with the wrong set of opinions, maybe.
While I'm sure there was a lot of attention on data security behind the scenes it is a little unnerving to hear someone say, when you aren't expecting it, "we looked at everyone's brokerage history and here's when they felt fear that they were about to lose everything."
I do wish there was a way to at least reduce the surprise factor. Perhaps put a note about the data collection protocol in the abstract on the article page.
Good security in one group does not mean good security in another group - at all.
Few people have any interest in finding out "who that person with cancer in that small town is", but "who is that unknown multi-millionaire in my parents' village" might be more tempting.
There's no political issue around individual health issues, but individual wealth is very political. I can definitely understand why someone wouldn't want identifiable information about their personal wealth given to third parties, especially not to US universities, which have a tendency to be rather left-wing and full of ideas of undeserved privilege that many don't share and would not feel comfortable knowing that their private information is shared with them. An angry research or student taking to Twitter to call out the evil capitalists isn't that crazy of an idea.
In other words, investors more worried about avoiding huge losses than they are about missing out, are more skittish. Potentially rational behavior, if your particular situation puts more emphasis on avoiding catastrophe, than on maximizing your chance of hitting it big long-term.
Not saying every instance of "freaking out" is rational, just pointing out that if you believe there's a 5% chance of losing almost everything, and a 95% chance of it being a buying opportunity, for some people that is a good reason to sell everything quick. Not everyone is in it for the long term, or most interested in maximizing their upside (rather than minimizing their downside risk). Age and dependents both point at investors who place greater emphasis on avoiding big losses than they do on attempting big wins, and they may be doing so for rational reasons.
Most of the conventional things that people understand about holding are wrong though...unsurprisingly, given that most of these conventions are produced by people trying to sell you something (I know this, because I worked in financial services, and used some of these conventions to sell things...it makes sense because individuals are stupid enough to do themselves real harm by trying to be clever...but this does not mean that being clever doesn't work, it just means you are stupid).
What do you mean?
If we flip this and imagine that everyone knew with certainty what the underlying values (earnings, dividends, etc.) should be then the return on stocks would always just be the return on the underlying values (it is a bit more complex than this, in reality the cost of capital differs between individuals so you are always going to get volatility...but the point is that because people are trade their portfolios faster than fundamentals change, then there is always going to be a timing strategy that works...that is how HFTs make money most days over decades).
In these cases, someone has lost most of their net worth and so they may find the rest of their assets are being liquidated, and they need to empty their account to discharge other liabilities. The reason why market downturns are correlated with panic selling is because market downturns are correlated with other things. The reason market cycles happen is largely because of the changing value that people place on cash...in 2008, lots of people needed cash.
And 90% is a very large number, the max drawdown on the S&P in crashes is usually less than half this number...so I am not quite sure why they are looking at this number when it doesn't really reflect a realistic scenario for most people. Losing 90% is inherently extreme, I worked in finance, I am relatively good at investing...if I lost 90% of my money, I would time myself out. I know of no investment business in the world that would keep someone on after a 90% loss...zero. Some multi-strategy firms will fire you if you lose 5%.
Additionally, as you say, people who have money tend to protect their downside. If you have $1m, and you lose 90%...you still have $100k. This tends to be particularly misleading when you look at a sample in which the market has always gone up and recovered...because it always looks like there was an opportunity cost from exiting (and if there wasn't, these papers don't get written). The paper is inferring reason to a decision that they don't really know the reason of. I 100% agree that "freaking out" can be (and often is) a rational choice.
Ask someone who "freaked out" in 1991 when Japan tanked. I actually know a firm that didn't "freak out", they were strong investors...firm nearly died, they lost pretty much all their client's money, and the Japan business effectively folded...but yes, those idiots who "freaked out"...rakakakaka.
This means that 45%-90% of equity was sold off ... they identifying the large sell-offs moreso than just equity drops.
Here is an example:
Month 0
STOCKS: $100
CASH: $0
Month 1
STOCKS $10
CASH $65
The account holder had their equity holding decrease by 90%. But that is composed of a market loss of $35 and their sale proceeds of $65.
65 / 90 = 72% (which is greater than 50% of the total decrease).
The 50% filter explicitly excludes those that just have market losses, the trader needs to have actively sold between 45-100% of their account value, combined with market losses between 0-45% for a total decrease of 90-100% between both factors.
I still have the same problem. The reason cycles cycle is because of the movement in money demand...which tend to be correlated to everything else. Being able to predict who will panic-sell could be correlated to other things (always a problem in financial economics).
I will need the read paper properly but would be interested to see how they deal with cash coming into and out of the account.
Someone should study that!
- .COM bubble
- Subprime mortgage crisis
- COVID
All three were nervewracking, but I held. In one case, I lost 101% on an Internet stock (CMTN for those who enjoy gloating). Total investments dropped 40% in the subprime crisis. COVID happened so fast, in both directions, that I really coudn't decide what to do.
In each case, not selling was the right thing to do. Yes, I could have done far better by selling early in the event and the buying at the bottom. But timing is not what investing is about.
This algorithm, if it's correct, can make a great deal of profit for the users: "major buying opportunity coming", at least until it's well known, at which point it might simply become a self-fulfiling prophecy.
Exactly. Markets are extremely efficient. As soon as this becomes known as reliable, markets will begin pricing these possibilities in as investors make decisions around it.
There is surely a study out there that demonstrates how each successive crisis has been shorter than the last. In the next crisis during which the fed has dry powder left, I fully expect the market to go directly up.
Mortgage crisis and COVID barely register. I love watching people freak out when stocks are down 3 or 4% in one day. It's my happy place. 1987 was 20% down in one day. If you're going to freak out about that, do something else with your money. It's disaster porn for me.
Then why do so few do it? It's because when there's blood in the streets and real fear it's for completely legitimate reasons.
What about sell when others are greedy?
This too is sage advice, except for one problem- when people are greedy nobody wants to sell, because the market is moving to ever higher prices for completely legitimate reasons.
The thing about trying to be contrarian is that true contrarianism isn't just fading what seems like a popular take. Rather, it's fading your own instincts when they reach extremes and telling you to make impulsive decisions, which is one of the most difficult hurdles for humans to overcome. We are incredibly influenced by our environments with major recency bias.
The truth about panics is that it isn't a real panic until everyone is panicking, bulls and bears because the reasons for the panic are completely real. Whether they prove to later be buying opportunities is totally irrelevant. The best traders who can actually not get caught up in the hoopla of a boom or bust are one in a million and still get it wrong more than half the time.
I have no dry powder to buy with. But I’m looking to change that this time, I feel a big multi decade level drop coming.
I may be wrong but I’ve got Warren Buffet and Michael Burry on my side.
Its not just instincts, if you extend the analogy, people dead or injured are not fighting. If a real estate speculator is underwater they aren't buying. Even if they want to buy they can't. Lots of people recognize that Miami/Vegas/Phoenix houses in 2010 were crazy cheap but people who weren't scared were broke.
If they were smart, they were selling on the way up.
The 1987 crash played out over about six weeks. The dot com bubble burst over the better part of a year. 1929 was about ten weeks from peak to the absolute bottom. But when the selling begins denial takes hold and everyone believes it’s an amazing buying opportunity. Most every crash has begun gradually then accelerated towards the end.
https://finance.yahoo.com/quote/%5EN225?p=^N225
What did machine learning add here? They seem to have found some straightforward correlations. Training a machine learning model just means you now have a model that you don't understand.
But you can sell it. With that model, you know who to cold-call.
Bottoms occur when the last people panic sell. People tend to extrapolate trends poorly.
Who do you think shorts are buying from? Panic sellers.
But anyway, in equities, which is what most people invest in, the longs greatly outnumber the shorts. So short covering has a much smaller impact than a reduction in panic selling.
Shorts usually cover once the rebound starts, not when things are free-falling (their trailing stops get triggered, etc)
Mar'20, 2008, etc - these did not happen because guys were getting margined. These happened because because were panic selling huge amounts of length.