Anchor Attention for Hybrid Crowd Forecasts Aggregation

Forecasting the future is a notoriously difficult task. To overcome this challenge, state-of-the-art forecasting platforms are "hybridized", they gather forecasts from a crowd of humans, as well as one or more machine models. However, an open challenge remains in how to optimally combine forecasts from these pools into a single forecast. We proposed anchor attention for this type of sequence summary problem. Each forecast is represented by a trainable embedding vector, and use computed anchor attention score as the combined weight. We evaluate our approach using data from real-world forecasting tournaments, and show that our method outperforms the current state-of-the-art aggregation approaches.