Probability of Informed Trading (PIN): Formula, Example, and Limits

Every completed trade has both a buyer and a seller, but the reason each side traded is usually not observed. A buyer may be responding to private information, rebalancing a portfolio, meeting a cash need, or executing a benchmark. PIN uses the counts of buyer- and seller-initiated trades over repeated periods to estimate how much order flow the model attributes to traders acting on private information.

Easley, Kiefer, O’Hara, and Paperman introduced the structural approach to study information-based trading and liquidity. Their model treats the daily order counts as outcomes of an unobserved information state. When private information is absent, uninformed buyers and sellers generate orders. When good or bad information arrives, an informed trader adds orders on the side implied by that information. The asymmetry in buy and sell counts across many days helps estimate the hidden state probabilities and arrival rates (Easley et al., 1996).

PIN is therefore an estimated property of a sample and a specified model. It is not a label attached to one print, a direct count of people with private information, or the probability that the next trade is informed. A high estimate says that the fitted model assigns a larger share of expected order arrivals to its informed-trader component than a lower estimate does, under the same definitions and assumptions.

The classic PIN model divides each trading period into one of three hidden states. With probability 1 − α, no information event arrives. With probability αδ, the event carries good news; with probability α(1 − δ), it carries bad news. The parameter α is the chance of an information event, while δ is the probability that an event is good news conditional on an event occurring.

Uninformed buy and sell arrivals each have an expected rate ε per period. An informed trader arrives at expected rate μ when an information event occurs and trades in the direction implied by the signal: buys on a good-news day and sells on a bad-news day. In the simplest version, buy and sell counts are conditionally independent Poisson variables within each state. The model does not observe the state or trader type; it infers their parameters from the resulting count patterns.

These are modeling assumptions, not descriptions known to hold for every market. They make the mixture model estimable and interpretable. They also mean that clustering, correlated liquidity demand, news announcements with different arrival patterns, trade-sign errors, or venue fragmentation can be mistaken for information-based order flow if the model does not account for them.

In the symmetric version, expected informed arrivals per period are αμ. Expected uninformed arrivals are ε from buyers plus ε from sellers, or 2ε. PIN is the informed component divided by total expected arrivals:

PIN = αμ / (αμ + 2ε)

The good-news parameter δ affects which side the informed trader uses, but it cancels from this aggregate share. Some model extensions allow more flexible buy and sell arrival rates or other trade-generating processes; their formulas and interpretations should not be substituted into the classic expression without explanation.

Consider a hypothetical period with α = 0.20, μ = 12 informed arrivals on an information-event day, and ε = 40 uninformed arrivals for each side per period. Expected informed arrivals are 0.20 × 12 = 2.4. Expected uninformed arrivals are 40 + 40 = 80. The model-implied share is 2.4 / (2.4 + 80) = 0.0291, or about 2.9%.

This example shows the arithmetic only. It does not mean that exactly 2.9% of observed trades were informed, that a trade has a 2.9% objective chance of being informed, or that any participant possessed inside information. The rates and event probability are hypothetical parameters; in real use they must be estimated from data and reported with the sample period and model specification.

For each period t, let Bₜ be buyer-initiated arrivals and Sₜ seller-initiated arrivals. Under the classic model, the likelihood for those counts is a three-part mixture. The no-event state has weight 1 − α and Poisson rates ε for both sides. The good-news state has weight αδ and rates ε + μ for buys and ε for sells. The bad-news state has weight α(1 − δ) and rates ε for buys and ε + μ for sells. Multiplying the corresponding Poisson probabilities within each state and summing across states gives that period’s likelihood.

The estimator chooses α, δ, μ, and ε to maximize the sum of the log likelihoods across the sample. PIN is then calculated from those fitted parameters. Researchers often estimate parameters over a window such as a quarter because a single day’s two counts contain little information about the mixture. The resulting PIN is a window-level estimate; increasing the window can stabilize counts while also averaging across changing market conditions.

The optimization is not a routine detail. The likelihood can have multiple local maxima, parameters can land on boundaries, and implementations can suffer floating-point underflow or overflow for large trade counts. Yan and Zhang document material bias from boundary solutions and propose a method for improving estimation (Yan and Zhang, 2012). A credible estimate should use a numerically stable likelihood, multiple defensible starting values, parameter constraints consistent with the model, and diagnostics that show whether the selected optimum is reproducible.

The classic model needs buyer- and seller-initiated counts. If a feed does not provide an aggressor-side field, those counts may come from a classifier such as a quote rule, tick test, or Lee–Ready procedure. Misclassified trades move observations between Bₜ and Sₜ and can change the likelihood. The classification method, treatment of midpoint and unknown trades, venue coverage, and timestamp alignment therefore belong in the estimation record. Trade direction classification explains why inferred signs can vary with the quote and the data feed.

The researcher must also define the period, instrument, eligible trade conditions, and handling of auctions, corrections, off-exchange reports, and no-trade intervals. Combining venues can create a different count process than analyzing one venue. Changing from daily to intraday periods changes the event-arrival interpretation and may leave too few observations for stable estimation. A PIN series is comparable across assets or time only when the data construction, estimator, window, and model version are sufficiently aligned.

Do not silently remove difficult observations or report only the estimate that looks plausible. Keep the original counts, document any exclusions, inspect parameter estimates and boundary hits, and compare the result under reasonable classification and window choices. If the estimate shifts materially, that sensitivity is part of the result rather than noise to hide.

A quoted or effective bid–ask spread measures a price concession or execution cost. PIN instead estimates the share of expected order arrivals attributed to the informed component of a structural model. A wide spread can reflect adverse-selection risk, inventory costs, fees, or competition; it is not itself a PIN value. The Roll spread estimator infers a spread from return autocovariance under its own assumptions.

Order-flow imbalance summarizes the difference in buy and sell activity, often alongside limit-order additions and cancellations. It can be observed directly over a chosen interval, while PIN uses a likelihood to infer a latent news process from repeated counts. An imbalance is not proof of informed trading: liquidity needs or clustered execution can produce a similar pattern. Order-flow imbalance describes that separate measure.

VPIN is also not simply a high-frequency PIN estimate. It is a volume-clock statistic built from imbalances in volume buckets and has distinct trade-classification and sampling choices. Andersen and Bondarenko’s assessment challenged claims that VPIN reliably measured toxicity or signaled turbulence in their tested setting (Andersen and Bondarenko, 2014). A researcher should evaluate each measure on its own construction and evidence rather than treat their names as interchangeable.

A fitted PIN value can look precise while the count model fits poorly. Gan, Wei, and Johnstone test the empirical fit of PIN models and report structural limitations in the distributions and dependence patterns they examine (Gan et al., 2017). This evidence is a reason to check fit for the actual sample, not a claim that every application is invalid.

Useful diagnostics include the distribution of buy and sell counts, their dependence, the frequency of zero-trade periods, the model-implied probabilities of those counts, the optimizer’s convergence and boundary status, and sensitivity to alternative starting values. Compare simulated or fitted count distributions with observed data and inspect whether parameters remain plausible across nearby windows. Standard errors or confidence intervals should reflect the estimation method; a single point estimate hides uncertainty from both sampling and model choice.

A stable estimate across adjacent windows is not proof that the model is correctly specified. Conversely, a changing estimate may reflect changing information arrival, a change in uninformed demand, data coverage, or misspecification. State the evidence and limits before assigning a market interpretation.

A reported PIN value is difficult to interpret without its measurement record. Identify the instrument or universe, venue coverage, sample dates, observation period and timezone, eligible trade conditions, and the procedure used to assign buyer- or seller-initiated signs. State whether the estimate uses the classic three-state model or an extension, and disclose the fitted α, δ, μ, and ε alongside PIN.

Also describe the estimation window, likelihood implementation, optimization routine and starting-value strategy. Report convergence, boundary solutions, uncertainty estimates, and the diagnostics used to assess fit. If auctions, corrections, no-trade intervals, or unknown trade signs were excluded or treated specially, say how. Those details let readers distinguish a robust pattern from a result driven by one data or optimizer choice.

The observation period matters because α is an event probability per period, while μ and ε are arrival rates per period. Changing from daily to hourly counts changes the units and the underlying event process; parameters from different period lengths are not directly comparable just because the PIN formula has the same shape. Align period definitions and trade construction before comparing assets or dates, then treat changes in the fitted rates as evidence to examine rather than automatically as changes in informed trading.

PIN was developed as a market-microstructure measure, not a rule for entering or exiting a position. It does not tell a trader which direction a price will move, how far it may move, whether an apparent signal is already reflected in the quote, or whether an execution can occur at a displayed price. The original asset-pricing study examined estimated PIN in a historical NYSE cross-section and related it to expected returns in that sample; its result is not a current forecast or a universal premium (Easley, Hvidkjaer, and O’Hara, 2002).

If PIN is used as an explanatory variable, report how trades were signed, what period was estimated, the estimation window, fitted parameters, uncertainty, model-fit checks, and sensitivity analyses. If it is considered in a trading strategy, test the complete rule on data not used to choose it, account for spread, fees, slippage, market impact, and capacity, and compare it with a baseline. Implementation shortfall provides one way to compare a decision price with actual execution costs. A statistically interesting estimate can still have no usable predictive or after-cost value.