Streak biases in decision making: data and a memory model

Streaks of past outcomes, for example of gains or losses in the stock market, are one source of information for a decision maker trying to predict the next outcome in the series. We examine how prediction biases based on streaks change as a function of length of the current streak. Participants experienced a sequence of 150 flips of a simulated coin. On the first of a streak of heads, participants showed positive recency, meaning that they predicted heads for the next outcome with a greater-than-baseline probability. As streak length increased, positive recency first decreased but then increased again, producing a quadratic trend. We explain these results in terms of outcome-prediction processes that are sensitive to the historical frequency of streak lengths and that make heuristic assumptions about changes in bias of the outcome-generating process (here, the coin). An ACT-R simulation captures the quadratic trend in positive recency, as well as the baseline heads bias, in two experimental conditions with different coin biases. We discuss our memory-based model in relation to a model from the domain of economics that posits explicit representation of an “urn” from which events are sampled without replacement.