Why Does Being Slightly Right, Many Times, Beat Being Brilliant Once?

In the early 1960s a young mathematician named Ed Thorp showed, with the help of an early IBM computer, that blackjack could be beaten. His method did not predict the next card. It tracked which cards had already been played, recognized the moments when the remaining deck tilted the odds slightly toward the player, and bet more in exactly those moments. The edge was often a percent or two.

Most people facing an uncertain decision reach for the opposite strategy. They look for the confident call: the hire who will transform the team, the campaign that will break through, the forecast that finally gets the quarter right. Judgment is measured by the hits, and the misses are blamed on bad luck. The approach feels decisive, but it stakes everything on a kind of foresight that almost nobody reliably has.

Thorp, and the systematic investors who followed him into markets, built a different discipline on a single reframing: the goal is not to be right, but to hold a small, genuine edge and apply it many times. That reframing sounds modest. Followed through, it reorganizes how evidence is gathered, how bets are sized, and how much trust any conclusion deserves.

Small edges need many bets

A probability is a more honest description of the future than a prediction. Saying a product has perhaps a 60 percent chance of launching on time invites planning for the other 40 percent; saying it will launch on time invites surprise. Once outcomes are described as odds, the natural question changes from "will this work?" to "is this a good bet, and how often will it be made?"

The second half of that question carries surprising weight. Renaissance Technologies' Medallion fund was reportedly right on only slightly more than half of its trades, yet it compiled one of the most remarkable records in investing. Richard Grinold and Ronald Kahn described why results like that are possible: performance grows with the skill behind each decision and with the number of genuinely independent decisions that skill touches. A small edge applied to thousands of trades, screening decisions, or pricing choices compounds into something large.

The same logic explains why written rules so often outperform experts on repeated judgments. Paul Meehl's 1954 review found that simple statistical formulas matched or beat clinical experts in most of the comparisons he examined. A rule does not get tired, anchored, or bored; it applies its small edge identically to every case, which is exactly what breadth requires.

Small edges look exactly like luck

Here the argument turns on itself. If the edges worth having are small, they are also faint, and a faint signal is almost impossible to tell apart from noise by inspection. Over a few dozen trials, a strategy that wins 51 percent of the time and one that wins 50 percent of the time produce nearly identical records.

That is why the hardest part of learning from data is not finding a pattern but proving the pattern is not luck. Give a model enough adjustable settings and it will fit any history perfectly, noise included. This is overfitting, and its signature is a strategy that looks brilliant on the past and fails on the future. The defense is to judge every idea on data it has never seen: a new quarter, a new region, a new cohort.

A subtler problem arises from volume: test enough ideas and some will look excellent by chance alone. Campbell Harvey and his co-authors, surveying the hundreds of return patterns proposed in academic finance, argued that the evidence required for a new "discovery" should rise with the number already tried. The consequence is uncomfortable: the more ideas tested, the higher the bar each winner must clear. A team that runs forty variants of a web page and celebrates the best one has quite possibly crowned an accident.

One further filter separates durable patterns from lucky ones. Persistent edges usually have a reason behind them: compensation for bearing a risk others avoid, a predictable human tendency, or a constraint that keeps competitors from acting. Asking who is on the other side of the bet, and why they accept it, is a statistical safeguard disguised as common sense. A pattern with a mechanism behind it is much harder to produce by accident than a correlation without one.

Tested edges must be combined and sized

An edge that survives testing is still only one bet, and one bet can fail. Harry Markowitz showed in 1952 that combining holdings that do not move in lockstep reduces the risk of the whole without proportionally reducing its expected return. The consequence reaches well beyond investing: the risk of a set of bets depends more on how they move together than on how risky each one is. Ten initiatives that share a single supplier, a single customer, or a single key person are one bet wearing ten labels.

Size matters as much as selection. John Kelly's 1956 formula, which Thorp carried from the card table into markets, sizes each bet in proportion to the edge and the odds. Because every edge is an estimate, careful practitioners bet a fraction of what the formula suggests, trading a little growth for protection against their own overconfidence. The principle generalizes to any budget of money, time, or attention: commit more where the evidence is strongest, and never so much that a mistaken estimate becomes fatal.

Every edge decays, and every model has boundaries

The last idea tempers all the others: an edge is worth only what survives its costs, and those costs include competitors. David McLean and Jeffrey Pontiff studied 97 published return predictors and found that their average returns fell by more than half after publication. Knowledge spreads, capital crowds in, and advantages erode. Every edge has a half-life, which turns research from a single discovery into a permanent pipeline.

Models also fail in ways their history never showed. Long-Term Capital Management, a fund whose partners included two Nobel laureates, collapsed in 1998 when markets moved into conditions its models had treated as remote: correlations that had been low rose together, and liquidity vanished. Benoit Mandelbrot had documented decades earlier that extreme price moves occur far more often than a normal curve predicts. A model is a tool with boundaries, and the discipline is to list the assumptions it depends on, watch them, and decide in advance when a person should overrule it.

The questions that remain

Taken together, these ideas form one method rather than a list of tips: describe the future as odds, and look for small edges that can be applied many times. Assume any promising pattern is luck until fresh data and a plausible mechanism say otherwise. Combine the survivors so that their failures do not coincide, size each according to the evidence, and count every cost. Then expect the edge to fade and the model to break, and keep building the next one.

None of this requires markets, equations, or code; it requires a handful of questions asked every time a decision repeats. What are the odds, how is that known, and could it be luck? What does it cost, how much should be committed, and where would the reasoning stop working? Anyone who asks them consistently is making many small bets in their own favor, and giving the arithmetic of repetition room to work.