AIrsenal decided to sell Erling Haaland: thoughts on minutes model, using TSB and price data

[rchan26]

Approximately at 9:30am on Friday 11th November 2022, AIrsenal made a move which no human would (akin to AlphaGo's "37th move") and made a transfer to sell Manchester City striker Erling Haaland (vs Brentford (H)) for Harry Kane (vs Leeds (H)) for gameweek 16.

At time of transfer, Haaland had 81.5% TSB compared to Kane's 25.7%. Whilst Kane is one of the best strikers the Premier League has seen (sitting in third all time at 192 goals), no "reasonable" FPL player would make this transfer.

There are a few potential reasons why this happened:

• By default, the prediction model looks back either the same number of weeks we're predicting for (here we just have one as it's the last week before the World Cup break where we have a wildcard to use). But it does look a minimum 3 weeks behind.
  • In the last three weeks (GW13-15), Haaland played 79, 0 and 26 minutes respectively due to injury. The model can perhaps be improved by improving the minutes model as per #486 to incorprate the injury data that we get from TransferMarkt
• Model does not take into account TSB (see #400)... should it?
  • Raises interesting questions about how "pure" we want the model to be, but IRL, players tend to keep players with high TSB as a way to "protect" themselves from losing out on points. With a player like Haaland with such high TSB, this is an extremely risky move, especially with how early this is in the season - there's not much consideration of "risk" in the model
• Algorithm optimises based on points, but it doesn't consider the fact that Haaland's price has increased over the time that he was in the team. Can we incorporate this information in the model? Is this useful information? Are there actual rules which use this that make sense?
  • Some FPL players try to accumulate money in the bank so that they can improve their team over time

[jack89roberts]

There's an old issue on your point about value here: #162