In 1973, psychologist Paul Slovic ran an experiment on eight professional horse race handicappers that has become a fixture in behavioural finance research. The design was simple. The implications for anyone who equates more research with better decisions are not.
The Setup
Slovic worked with men who handicapped races for a living, not casual bettors. He stripped horse names from real historical race data so reputation and memory could not contaminate the test, then asked each handicapper to rank a large pool of possible variables by importance: recent form, jockey experience, weight carried, and dozens of others.
The test itself ran in four rounds. Round one gave each handicapper only the five variables he had rated most useful. Round two gave him ten. Then twenty. Then forty. After each round he picked winners and stated his confidence as a percentage.
Slovic wasn’t trying to find the best handicapper. He was asking what happens to judgment as information increases.
Accuracy Stalls, Confidence Doesn’t
With five variables, the handicappers picked winners at roughly 17 percent accuracy, comfortably ahead of chance. That much makes sense: some information is better than none.
What happened next is the part that gets quoted endlessly. As the handicappers moved to ten, twenty, and forty variables, accuracy stayed essentially flat, statistically indistinguishable from the five-variable round. Eight times the information bought them nothing in predictive power.
Confidence moved in the opposite direction. By the fortieth variable, average stated confidence had roughly doubled, to around 34 percent, despite no corresponding gain in accuracy. The extra data didn’t make these professionals better at their jobs. It made them feel better at their jobs, and that gap between feeling and fact is the mechanism worth paying attention to.
Carry that mechanism into a brokerage account and the pattern is familiar. An investor who has read the earnings call transcript, five years of filings, three sell-side notes, and a dozen forum threads isn’t necessarily more accurate. He is almost certainly more confident, and that confidence is what sets position size, whether a stop-loss gets used, and whether disconfirming evidence gets a hearing once he’s already in the trade. Overconfidence rarely causes losses directly. It causes oversized bets, and oversized bets are what turn an ordinary miss into a serious one.
Not Everyone Got Worse the Same Way
The detail most summaries skip is more useful than the headline number. The eight handicappers didn’t move together. Three actually got less accurate as they received more variables. Two improved. The rest stayed flat. The averaged result, flat accuracy against rising confidence, hides real variation in how individuals handled the extra data.
A parallel result shows up in a very different domain. In 2017, Sendhil Mullainathan and coauthors examined 554,689 New York City defendants from 2008 to 2013, of whom human judges had released about 400,000 pending trial. The researchers built a machine learning model, fed it a narrow slice of the same information, and had it generate its own release list, then checked both lists against who actually reoffended before trial.
The comparison wasn’t close. Defendants the algorithm selected for release were 25 percent less likely to commit a crime while awaiting trial than those judges released; roughly one in four people judges released went on to reoffend. Among the small slice of defendants the model flagged as high risk, more than 50 percent likely to reoffend, judges still released 48.5 percent of them. Across the full dataset, judges released more than 2,600 high-risk defendants the model would never have considered releasing.
The asymmetry in inputs is the interesting part. Judges had the defendant’s full record, testimony from prosecution and defense, and everything they could observe in person, demeanor, remorse, a spouse in tears. The model had two inputs: age and prior record. It didn’t see any of the human context, and it still made better predictions. The additional information available to judges didn’t sharpen their judgment. It clouded it.
The Real Lesson Isn’t “Less Data”
The honest takeaway isn’t that less information is always better. Two of Slovic’s eight handicappers genuinely improved with more variables. Most didn’t, and a few were actively harmed, likely because they began treating noisy, low-value variables as if they carried real signal. Later analysis of the same data found that experts tended to overrate the importance of low-predictive-value factors precisely when they had more of them to choose from. Give someone forty inputs and they generally won’t discard the useless ones; they’ll build a narrative in which most of them feel load-bearing.
For an investor, this reframes the problem. It isn’t about how much data you’re looking at. It’s whether you can honestly separate the inputs that have ever changed a call for the better from the ones that just make the process feel thorough. Most investors have never actually tested this about themselves. They add more research the way the handicappers requested more variables: because it feels responsible, not because anyone verified that the tenth data point ever moved a decision.
The practical fix isn’t to cut your process to five inputs and stop. It’s auditing, on a regular basis, which parts of your process are actually earning their place. That means writing down, before a decision, exactly which inputs you used, then later checking whether the outcome would have changed had any specific input been excluded. Almost nobody does this, because it’s uncomfortable, and because the alternative, just gathering more, feels like diligence even when it functions as a confidence-generating ritual with no measurable payoff.
Slovic’s handicappers weren’t unusually foolish. They were working professionals with genuine expertise, which is exactly what makes the result hold up. Expertise didn’t close the gap between feeling right and being right. There’s no strong reason to think it will close that gap for anyone else, including an investor convinced that one more report will finally tip the odds in his favour.
