People made forecasts from real data series. The points in the series were un‐trended and independent. Hence, forecasts should have been on the mean value. However, consistent with previous research on forecasting biases, forecasts were too close to the last data point. It appears that forecasters see positive sequential dependence where none exists. In three experiments, we examined this bias in different types of forecasting task: point forecasting, probability density forecasting, and interval forecasting. In all cases, we found that it was greater when the data series were displayed using continuous line graphs than when it was displayed using discrete point graphs. Consistent with arguments made by Zacks and Tversky (Memory and Cognition, 27:1073, 1999), we suggest that people are more likely to group data together and to see patterns in them when those data are presented in a continuous than in a discrete format. These findings have implications for forecasting practice.

When people make forecasts from series of data, how does their accuracy depend on the length of the series? Previous research has produced highly conflicting findings: some work shows accuracy increases with more data; other research shows that it decreases. In two experiments, we found an inverted U-shaped relation between forecast error and series length for various series containing different patterns and noise levels: error decreased as the length of the series increased from five through 20 to 40 items but also decreased as the series length decreased from five through two to one item. We argue that, with short series, people use a simple heuristic approach to forecasting (e.g., the naïve forecast). With longer series, they extract patterns from the series and extrapolate from them to produce their forecasts. Use of heuristics is poorer but extraction of patterns is better when there are more items in the series. For series of intermediate length, neither type of strategy operates well, thereby producing the inverted U-shaped relation that we observed. Implications for unaided judgmental forecasting and for forecasting based on a combination of judgmental and statistical methods are discussed.

In two experiments, forecasters made a sequence of five forecasts from different types of time series, either from the nearest horizon to the most distant one (1, 2, 3, 4, 5) or in one of two other orders, both of which required the forecast for the most distant horizon to be made first ('endanchoring'). These latter two orders differed in terms of the direction of the remaining forecasts: it was a horizon-increasing order (1, 2, 3, 4) or a horizon-decreasing one (4, 3, 2, 1). End-anchoring improved forecast accuracy, especially for more distant horizons. It resulted in the trajectory of the forecast sequence being closer to the optimal one. Direction of forecasting after end-anchoring affected forecast quality only when the optimal trajectory of the forecast sequence displayed strong nonlinearity. End-anchoring provides a simple means of enhancing judgmental forecasts when predictions for a number of horizons are made from each series.

Wave forecasting is accomplished today via numerical models. In this work we apply stochastic techniques using actual measurements to improve wave height forecast in real time. Application of these techniques in four locations of the Aegean Sea results in significant improvement of the forecast in the time domain retaining the same pattern of modifications, suggesting, thus, this method for operational use in deep and intermediate waters. The improvement is obtained by four regression models, which take into account the variable of the significant wave height as measured and forecasted by the model. Space-wise extension of the method was also investigated and applied to the Aegean Sea and the Indian Ocean, where its performance was remarkable.