When Everyone Sees the Same Signal: How a Good Prediction Becomes a Bad Trade

A signal is worth what the people who do not have it are willing to pay for it. That is an uncomfortable premise for a research organization, because the usual measures of a signal's quality, its information coefficient, its stability across subperiods, its robustness to small changes in specification, say nothing about how many other participants are computing the same thing from the same data. Two firms can each run a careful backtest on a public dataset with a textbook method and each conclude, correctly by every internal standard, that they hold an independent edge. They do not. They hold the same edge, and they find this out the first time both try to trade it in size.

This piece is about what happens to expected return as that discovery is made across a market. The process is usually called crowding, and we think it is best understood as two distinct kinds of decay that are often run together. The first is signal decay: the forecast becomes less accurate because the prices it is applied to already reflect it. The second is execution decay: the forecast remains accurate, but the trade that acts on it becomes expensive because the participants who share the forecast also share its timing. The two have different signatures, different remedies, and different points at which they turn a good prediction into a bad trade.

Two Ways to Lose the Same Return

Begin with a single mispricing. A signal identifies a security whose price sits some distance from where the signal expects it to be a month from now. The expected return to holding the security is that distance, and the information coefficient, the correlation between the signal's rankings and subsequent returns, measures how reliably the signal finds such distances across many securities. Now let the capital positioned on the signal grow. Each participant who acts on it buys the security and moves the price part of the way toward the forecast. The next participant computes the same forecast from a price that has already moved, and the distance that remains is smaller. The signal has not become wrong. It has become priced, and the information coefficient measured against current prices falls accordingly.

Figure 1 draws that relationship in a stylized model. The horizontal axis is a crowding index: capital positioned on the signal, expressed as a multiple of the amount at which the uncrowded edge would be halved. The vertical axis is the information coefficient. We draw three versions of the curve because the shape depends on how quickly prices absorb positioned capital, and that speed is not something a backtest can reveal. The dashed line is what the backtest reports, because the backtest was run on prices from before the crowd arrived.

The dashed line deserves a closer look. A backtest does not report the information coefficient of a signal; it reports the coefficient at the level of crowding that prevailed during the sample. If the sample predates the signal's publication, the backtest describes a market that no longer exists. That is why a signal can be found, validated with every reasonable safeguard against overfitting, and still disappoint out of sample without any error in the research. The research was right about a market with fewer participants in it.^[1]

Where the Return Goes

Execution decay is different in kind. It arises not from the price at which the signal is computed but from the price at which the resulting trade is filled. Many signals fire for many participants at the same moment: the same data release, the same month-end rebalance, the same close. When they do, everyone buys the same names in the same window, and the cost of demanding liquidity at that moment depends on the total flow rather than on any one participant's share of it. In the classic model of a market maker facing anonymous order flow, price impact is proportional to the net quantity demanded, and the market maker cannot tell whose orders are whose.^[2] Each participant therefore pays for the crowd's flow as if it were its own. The forecast is still right at the decision price. It is simply no longer available at that price.

Figure 2 puts the two effects together in a stylized decomposition. Each bar is the uncrowded expected return of a signal, held at an illustrative 40 basis points per period, split into three pieces at increasing levels of crowding: the part lost because the forecast has already been priced, the part lost because the fill is worse than the decision price, and the part the portfolio keeps. The signal-decay piece uses the proportional curve from Figure 1; the execution-decay piece adds a linear impact cost.

Two features of the picture carry over to practice. First, signal decay is the larger piece early on, but it is concave: each additional unit of capital removes less of the forecast than the last, because there is less left to remove. Execution decay is smaller at first and does not flatten; in this model it keeps growing in proportion to the crowd. By the time positioned capital reaches the half-edge level, the forecast is still half as good as the backtest said and the trade keeps nothing. Second, the two pieces answer to different remedies. Signal decay can be addressed only by having a different forecast: different data, a different horizon, a different way of combining the same inputs. Execution decay can be addressed by trading at a different time from the crowd, by spreading the trade over a longer window, or by holding a smaller position. One is a research problem and the other is an execution problem, and a firm that treats a crowded signal as a single problem will apply the wrong fix to at least half of it.

The Break-Even Point

The question a portfolio needs answered is not whether a signal is decaying but at what size the trade stops being worth making. The natural way to frame it is per dollar. As capital positioned on the signal grows, the expected edge on each additional dollar falls, for the reason Figure 1 describes; at the same time the impact cost of each additional dollar rises, for the reason Figure 2 describes. The point at which the two schedules cross is the break-even, and it is a property of the signal in its crowded state, not of the signal in isolation.

Figure 3 draws both schedules against capital positioned on the signal, on a logarithmic scale, in the same stylized model. The dollar figures are arbitrary and chosen for legibility; the shape is what matters. Expected edge per dollar starts at the uncrowded 40 basis points and halves at a reference level of capital. Impact cost per dollar rises in proportion to the capital being deployed alongside. The net line is the difference, and the marker sits where it crosses zero.

To the left of the marker the prediction is a trade. To the right it is still a prediction, and a reasonably good one, but no longer a trade. In this model the crossing arrives when the edge is exactly half what the backtest reported, which is the point to hold on to: a forecast is a long way from useless at the moment the trade built on it becomes worthless. A concave cost schedule, in which impact grows with the square root of flow rather than in proportion to it, moves the crossing to the right but does not remove it, and a firm that assumes the concave schedule when the linear one applies will size past the crossing without knowing it.^[3]

The practical corollary is that the capital positioned by others is an input to the sizing decision, whether or not one has a good estimate of it. It cannot be observed directly, but it leaves traces: in the correlation of a signal's returns with the returns of other systematic participants, in the coincidence of flows around known rebalance dates, and in the cost of trading in the signal's own direction relative to the cost of trading against it. A sizing rule that ignores these is implicitly assuming a crowding index of zero, which is the one value it is almost certainly not.

Correlation Is the Symptom

The uncomfortable thing about crowding is that very little of the above is visible in the signal's own returns until the break-even has been crossed. What shows up earlier, and in a different place, is correlation. The signal families a systematic investor combines, valuation, momentum, quality, earnings revisions, short-term reversal, low volatility, are constructed to be close to independent bets. That independence is a statement about the signals. Crowding adds a common factor that is a statement about their owners: the flow of capital into and out of systematic strategies as a class. When that capital withdraws at once, because a drawdown at one participant forces deleveraging or a risk limit tightens across many, signals that were independent in construction fall together, because the same participants hold them. The episode of August 2007 is the canonical illustration: strategies designed to be uncorrelated declined together over a few days and recovered together, with little in the underlying economics to explain either move.^[4]

Figure 4 shows the starting point in a one-factor model of six stylized signal families. Before crowding, each signal loads lightly on a common market-structure factor and otherwise on its own idiosyncratic component, so the pairwise correlations are small. The matrix is close to what the signals' designers intended.

Figure 5 repeats the matrix after crowding. Every signal now also loads on a shared flow factor, the capital that moves into and out of all of them at once, and the correlation matrix fills in. Nothing about the forecasts has changed between the two panels; the loadings on the market-structure factor are identical. The signals have not changed. Their owners have.

The consequence for portfolio construction is direct. A portfolio sized on the first matrix is levered to the risk in the second. Worse, the second matrix is state-dependent: the shared flow factor is quiet most of the time and loud precisely when the crowd exits, so an estimate of correlation from ordinary periods recovers something close to the first matrix even after crowding has occurred. Diversification across signals is only as real as the independence of the people who hold them, and that independence has to be estimated, not assumed from the way the signals were built.

What We Watch, and What We Do

None of the quantities above can be observed cleanly, but each leaves a trace that can be monitored, and the traces differ by kind of decay. Figure 6 collects the ones we find most useful. The point of the table is the mapping in its middle column: the same fall in realized return has a different diagnosis, and a different response, depending on which symptom accompanies it.

Three practices follow. The first is to measure the information coefficient twice, once against decision prices and once against fill prices, and to track the gap between them as a series in its own right. A widening gap with a stable decision-price coefficient is execution decay and calls for a change in trading, not in research. A falling decision-price coefficient is signal decay and calls for the opposite. The second is to size on the crowded edge rather than the backtest edge: to treat the capital positioned by others as a parameter to be estimated, to estimate it from the traces in Figure 6, and to let the estimate move the position toward the break-even rather than past it. The third is to treat correlation with the crowd as a risk factor in construction, so that a portfolio of signals that were independent on paper is not unknowingly a single bet on the patience of other systematic investors.

Underneath all three is a preference for the kind of edge that is expensive to share. Data that is difficult to assemble, a horizon that others find inconvenient, an execution capability that lets a forecast be traded when the crowd is not trading it: none of these makes a signal's prediction better. They make the prediction one's own for longer, and in a market where everyone eventually sees the same signal, that is the part left to compete on.