Mr. Tippy's Tips

Mr. Tippy receives tips from real users in Slack, giving us a unique opportunity to explore AFL match predictions through a wisdom-of-the-crowd model. Wisdom-of-the-crowd operates under the assumption that as more tippers participate in Mr. Tippy competitions, their aggregate predictions become increasingly capable of making accurate tips. While the performance of our model could possibly be improved by blending users' tips with other meaningful signals — such as weighting tips by tippers' performance — we have chosen a model we think best represents the simplicity of letting the crowd's wisdom speak for itself.

Confidence

For each match, Mr. Tippy predicts whichever team is tipped by the majority of real tippers in Slack (auto-tips excluded). The confidence value represents the percentage of tippers who picked the favoured team — so if the split is close, the confidence value will reflect that in a value close to 50%. In the case of a tie, confidence is shown as 50% and the predicted winner is taken from the highest-ranked tipster across all competitions.

The confidence value for each game of the current round can be seen in the table above. Confidence values tend to be high — sometimes even 100% — revealing substantial differences between how algorithmic models calculate confidence and how it is expressed through the choices of real tippers. While this value is not itself used to tip in Slack, it is integral to the calculation of both bits and Mean Absolute Error (MAE) metrics used to compare our wisdom-of-the-crowd model to Squiggle-approved models.

Margin — First Game of Each Round

Mr. Tippy makes two types of margin predictions, starting with the first game of each round just like real tippers do in Slack. For this game, Mr. Tippy takes the margin predictions from tippers who have picked the same team as the majority, and then identifies the median of these predictions. This prediction is based only on real tips and can be seen in the table above next to the prediction of the first game. It is used by Mr. Tippy to tip in Slack.

Margin — All Other Games

Tippers in Slack only make margin predictions on the first game of each round. However, Mr. Tippy needs to calculate margins for every match in order to compare its performance against other AFL prediction models (see table below). To do this, we have to extrapolate margins from binary tips.

The most natural starting point here is to use our confidence value as an indicator of how close games will be. As we have already explained though, our confidence values tend to be quite high in comparison to algorithmic models, and so the relationship between confidence and margin is difficult to determine.

For this reason, and to allow the crowd's wisdom to speak with minimal intervention, we use an ordinary least squares regression to calculate margins for all other games. Predictions are derived from the crowd's confidence (p, the percentage of tippers that picked the favoured team) using a simple linear formula: margin = a × p + b.

The values of a and b are not fixed. They start each season at a = 135, b = −72 — values chosen to closely mimic the logistic curve margin = 24 × ln(p / (1 − p)) — and are then refitted after each completed round using ordinary least squares regression on all prior correct predictions. This means the formula continuously adapts to the accuracy of real crowd confidence as the season unfolds. The final margin is clamped to be at least one point, so that our model never predicts negative margins that contradict the match pick.

Through using a simple adaptive equation, we aim to present a model that preserves the spirit of wisdom-of-the-crowd predictions, while also helping uncover the relationship between crowd vote share and margin results.