A Million Small Bets: Breadth, Skill, and the Fine Print of the Fundamental Law

A forecaster who is right 52 percent of the time is nearly useless on any single call and nearly unstoppable across ten thousand of them. That sentence is the whole argument for systematic investing in miniature, and like most short arguments it is true only under conditions that deserve to be spelled out. Those conditions are the subject of this piece. A quantitative portfolio is not a collection of strong views. It is a machine for making a very large number of weak ones, each too small to matter and too numerous to ignore, and its robustness comes from the arithmetic of their sum rather than from the quality of any one.

The arithmetic, which we call here simply the law, fits on an index card. We set it out below, then spend most of our time on what it leaves out: what actually counts as a bet, why bets that look independent rarely are, why skill measured on the first hundred ideas seldom survives to the thousandth, and why the number in the title is an aspiration about counting rather than a description of any real portfolio.

The Law, Stated Carefully

The law says that the expected information ratio of an active strategy, its expected excess return per unit of active risk, is approximately the product of three terms: IR ≈ IC × √BR × TC. The information coefficient, IC, is the correlation between forecasts and subsequent outcomes, a measure of skill per bet. Breadth, BR, is the number of independent bets the strategy makes per year. The transfer coefficient, TC, is the correlation between the ideal portfolio the forecasts imply and the one actually held after constraints and costs have done their work; it is the fraction of the forecast that survives implementation.

Two features of the formula do most of the work in what follows. The first is that skill enters linearly while breadth enters through a square root: doubling the IC doubles the expected ratio, and doubling the number of bets raises it by about 41 percent. The second is that the three terms multiply, so a weakness in any one of them scales the whole product. A strategy with unusual skill and no breadth, or enormous breadth and no way to implement it, is not rescued by its strength elsewhere.

Figure 1 draws the formula for three levels of skill, with breadth on a logarithmic axis, and adds one dashed curve to show what a transfer coefficient of one half does to the strongest of them. The ICs are hypothetical and, by the standards of most real forecasts, generous. The point of the picture is the shape, not the level. On the logarithmic axis the curves bend upward, because each tenfold increase in breadth multiplies the ratio by the same factor, a little over three. Read against N itself the same curves are concave everywhere: each doubling of breadth adds the same 41 percent, so each additional bet is worth less than the one before, yet the curves never flatten, so there is no number of bets beyond which breadth stops helping. The upper right of the chart, where the model produces ratios that no one should expect to observe, is a region the rest of this piece explains how not to reach.

It is worth pausing on how modest the skill in Figure 1 is. An IC of 0.02 means that a forecast explains a few hundredths of one percent of the variance of the thing it forecasts. No one would act on such a forecast in isolation. The law says that acting on it a thousand times a year, independently, produces an expected ratio comparable to what a handful of very good calls would produce, and that acting on it ten thousand times produces something better. That is the sense in which breadth substitutes for conviction, and it is the reason a systematic firm can be indifferent to the fate of any single position.

What a Year Looks Like

The law is a statement about expectations. What a portfolio's owner experiences is a sequence of realized years, and it is in the realized years that the case for breadth becomes visceral. Consider a stylized bettor with a per-bet edge of four percentage points, right 52 percent of the time and wrong 48, who sizes each bet so that the year's total risk is the same however many bets are made. With ten bets a year the expected result is about an eighth of a unit of risk, and the realized result is almost entirely noise: the bettor loses money in something like 45 percent of years, and two losing years in a row happen about one time in five. With a thousand bets the expected result is ten times larger, a little more than one unit of risk, while the noise is unchanged, and a losing year becomes roughly a one-in-ten event.

Figure 2 shows the two distributions side by side, in the same units. The width of the two histograms is identical by construction, because total risk has been held fixed; what breadth changes is where the distribution sits relative to zero. Read the other way, with every bet sized identically instead, the same arithmetic appears as a tenfold narrowing of the diversified book's outcomes around its expectation. Either reading is correct. The concentrated bettor's year is dominated by luck; the diversified bettor's year is dominated by the edge.

This is the practical meaning of robustness. A robust portfolio is not one that avoids bad outcomes, since no portfolio with risk in it does that. It is one whose realized results track its expected results closely enough that a bad year is informative rather than merely unlucky. When a thousand-bet book loses money, something about the forecasts is probably wrong, and the process can be examined. When a ten-bet book loses money, nothing has been learned, because the same process would have lost money 45 times in a hundred regardless.

What Counts as a Bet

Everything so far assumes the bets are independent, and this is the assumption that fails most visibly. Breadth in the law is not the number of positions, or the number of trades, or the number of securities in the universe. It is the number of independent forecasts, and independence is scarce. A portfolio of two thousand stocks tilted toward a single characteristic, cheapness, say, or recent price strength, is making one bet two thousand times. Its positions are numerous and its breadth is close to one. The realized outcome of that portfolio will look like the left panel of Figure 2 however long its holdings list, because the two thousand positions succeed and fail together.

The same is true across time, a point we return to below: a forecast that barely changes from one month to the next is the same bet held longer, not a new one.

The arithmetic of correlated bets is simple and unforgiving. If N bets share an average pairwise correlation ρ, the number of independent bets they are worth, their effective breadth, is N divided by 1 + (N − 1)ρ, and as N grows that expression approaches 1/ρ. Figure 3 draws it. The dashed diagonal is the frictionless case, in which every bet counts. The three solid curves peel away from it one after another and flatten, each at the ceiling its correlation implies: a correlation of only 0.01 among bets caps their effective number at one hundred, no matter how many are made. A million nominal bets with a pairwise correlation of one thousandth are worth about a thousand independent ones. The title of this piece describes a counting exercise, not a portfolio.

Two consequences follow. First, the marginal bet's value depends less on its own skill than on its correlation with the bets already in the book. A mediocre forecast that is uncorrelated with everything else can raise effective breadth more than a strong one that duplicates an existing exposure, which is why research effort in a systematic firm is so often directed at finding different information rather than better information. Second, the correlation among bets is not a constant of nature. It rises in stressed markets, when securities begin to move together for reasons that have nothing to do with the forecasts, and so effective breadth is lowest precisely when it is most needed. A portfolio that reports its breadth as a single number is reporting an average over regimes that are not alike.

Where the Law Bends

The law is an approximation with a list of assumptions attached, and being honest about it means being explicit about where each one bends. Figure 4 collects the ones that matter most in practice. The most consequential is the assumption of constant skill. The IC in the formula is one number, applied to every bet; in a real research process it is an average over bets of very different quality, and the average falls as the process reaches further for breadth. The first hundred ideas in a firm's inventory are the ones it understands best. The thousandth is there because the first hundred were not enough, and it is usually weaker, noisier, and more likely to be a statistical accident. Breadth pursued carelessly does not add independent skill; it dilutes the skill that was already there.

The second assumption is implementation. The transfer coefficient is a modest-looking term, but it multiplies everything, and it is the term most under a firm's control. Limits on position size, leverage, and turnover, the cost of trading, and the delay between a forecast and the trade that expresses it all lower it. Breadth makes this harder, not easier: the fastest way to add independent bets is to refresh forecasts more often, and every refresh is a trade with a cost attached. A strategy can add breadth on paper and give it back in the market.

The third is stability. The IC and the correlation structure that go into the formula are estimates from history, and both drift. A firm that computes its expected information ratio from last decade's IC and last decade's correlations is describing last decade. The law says what to expect if the inputs hold; it does not say whether they will.

Bets Across Time

Breadth has a second dimension that is easier to inflate than the first: time. A forecast refreshed every month looks like twelve bets a year. Whether it is depends on how much the forecast actually changes between refreshes. A slow signal that says nearly the same thing in March as it said in February has not placed a new bet in March; it has repeated an old one and paid to trade it again.

A stylized example makes the size of the gap concrete. Suppose a monthly forecast carries most of its value from one month into the next, so that consecutive readings have a correlation of 0.9. Under a simple persistence model, twelve such readings carry roughly the information of one fresh, independent reading, not twelve. The arithmetic is illustrative, not a measurement of any Oak St. signal, but the direction is general: the more persistent the forecast, the fewer bets the calendar actually contains.

Our rule follows from it. We count a bet when the information changes, not when the trade happens. A refresh that moves the forecast is breadth; a refresh that only restates it is turnover, and turnover is a cost. This matters most for any programme built on a small number of deep, exchange-traded markets, where breadth across instruments is limited by design. In that setting the honest move is not to manufacture breadth by trading more often, but to admit that the book holds a handful of real bets and to size each of them as what it is.

How We Think About It

Figure 5 gathers the arithmetic into four numbers, and they are the numbers we keep in front of us when deciding what to build. At the margin, skill is worth more than breadth, in the sense that doubling one doubles the result while doubling the other adds 41 percent; but skill is also far harder to double, and so breadth is where most of the practical leverage lies. Correlation caps breadth near 1/ρ, so the useful question about a new forecast is rarely how good it is and usually how different it is. The transfer coefficient multiplies everything, so implementation is not a downstream detail but an equal part of the design. And the model's expectations are only as good as its inputs.

Several practices follow. We count bets in effective rather than nominal terms, estimating the correlation structure of forecasts as carefully as their skill and reporting breadth as a range across regimes rather than a point. We value a candidate forecast by its expected contribution to effective breadth, which means a new signal is judged first on what it adds that the existing inventory does not already contain. We treat the transfer coefficient as a design variable, so that constraints, trading costs, and the speed of implementation are argued about at the same table as the forecasts themselves. And we hold the skill estimate for any single forecast loosely, because the law's promise is only as good as the IC, and the IC is the term most easily flattered by a backtest.

None of this diminishes the law. It is a statement about the shape of the problem, and the shape is right: a very large number of small, weakly correlated forecasts, honestly counted and efficiently implemented, is a more robust foundation for a portfolio than a small number of large ones, not because the small forecasts are better, but because their sum is more predictable than any of its parts. A million small bets is the ideal. The work is in the counting.