Thread by @SebWassell

I think there is a misunderstanding or miscommunication over what an #FPL model is telling us when it makes a points 'prediction'.

Chiefly, the fact that it isn't really a "prediction" but a projection:

🧵

In any given gameweek, a player will have a projected number of points. These are summed over a period of gameweeks for a longer term projection.

For example, from #GW14 to #GW16 these are the points projections for two players, A and B, from @fplreview's Free Model:

However, this does not necessarily mean that the model thinks Player A will score 6.6 pts in GW14 or 15.6 pts from GWs 14-16.

It means that the average return from 1,000s of simulated runs of GW14, and by extension the sum of GWs 14-16, is 6.6 and 15.6 respectively.

These simulated runs settle into probability distributions, usually 'normal' probability distributions:

Given 1,000s of goes at a single gameweek (with the same inputs*) the model will produce a range of potential results, from the extremes (eg. scoring 5 goals or getting sent off in the first half after missing a penalty) to the less extreme (eg. 60+ minutes and an assist).

Instinctively we know all of these outcomes are possible, even if many are not probable. A model running 1,000s of simulations can attach a numerical value to these probabilities.

The more probable outcomes naturally fall into the middle of this pattern, closer to the mean.

It is this mean that is presented to the user of the model, not as the single most likely outcome, but the average of 1,000s of possible outcomes.

In a simulation run 3 times, a mean of 6.6 points could be derived from 9 pts + 9 pts + 2 pts (20/3 = ~6.6). But the mode is 9, etc.

It can be tempting to take this mean as a points prediction and claim a failure in the model if the reality is very different.

However, the reality was almost certainly captured in the probability distribution, it simply wasn't a particularly likely outcome. But that's football!

A further hurdle arises when we consider how points are accumulated in FPL.

I'll use the above example, where Player B is projected to score 5.0 points in GW16.

5.0 points could be accumulated in a number of ways.

Let's say a midfielder accumulates 60+ minutes (2), no cleansheet (0), one assist (3) and no goals (0), bonus (0) or any other event that gives/takes points.

The model would be 'correct' in predicting 5.0 points.

However, only a single, potentially fleeting, event - such as laying the ball off to another player 50 yards from goal who goes on to score a worldie - could almost double the actual outcome (5 + 3 [assist] = 8), even before we consider bonus ramifications.

As such, an extremely marginal real world event has had a huge impact on the points and by extension the perceived accuracy of the 'prediction', when in reality the projection was pretty good given the probabilities it assigned to various potential outcomes.

Thinking of that prediction as a projection, where we simply got pushed a little further away from the average and closer to an extreme, is a more robust way of considering the efficacy of models and how you should use the information they present.

When then comparing two players in this way, it is perhaps better not to compare the literal projections themselves and say that 15.6 < 14.1 therefore Player A wins, but rather how confident a 1.5 point difference over three weeks should make us in saying saying that A > B.

*It is entirely fair to interrogate the inputs used to formulate these projections, but that is probably one for another time!

I've just realised that I may have accidentally used the different averages interchangeably here, when accuracy in that department is quite important.

I'll tag @fplreview for clarity. There is a chance that "mean" should be 'median'.

The point remains unchanged by this however!