R19 Captain

Cook, Haas or Smith the obvious contenders

Cook!!!

Ben Marlin wrote:Cook!!!

I worry that your model lacks depth and controls, but I love that you're running this sort of analysis.

Cook+1 from me.

I love it too, however, the big problem from the analysis that I can see is the point that Milchy kinda raised and that is that everyones scoring has picked up (on average) over the season. Therefore any model that focuses on one player in isolation with a weighting toward time series analysis is gonna tend to predict above average performance this week.

I have concerns about the time weighted part of the prediction
The part of Marlin's analysis that indicates Dragons often concede big points to hookers I think is much more useful.

Cook probably would have captain anyway, take out origin period and he's been a great reliable captain

Yeah Cook was the set and forget captain until origin.

Even Verbal said prior to trading Cook out that one thing he learned from last year was (C)ook and it was working at that stage.

L-Jimmy wrote:

I  worry that your model lacks depth and controls, but I love that you're running this sort of analysis.

Cook+1 from me.

Thanks mate
How do you suggest that I could add depth and controls?

Ben Marlin wrote:
Thanks mate
How do you suggest that I could add depth and controls?

Good question and L Jimmy is the one to ask.

Can I just ask, Are u using OLS to estimate the equation? Also if u don’t mind me asking are u a student?

Also , I really hope u are not taking offense to any of the feedback myself and others are giving. I think what u r doing is great and adding a lot to the forum. The fact that u have people like milchy l jimmy etc commenting is testament to that. So good stuff man.

Rippin and Tearin wrote:
Good question and L Jimmy is the one to ask.

Can I just ask, Are u using OLS to estimate the equation? Also if u don’t mind me asking are u a student?

Also , I really hope u are not taking offense to any of the feedback myself and others are giving. I think what u r doing is great and adding a lot to the forum. The fact that u have people like milchy l jimmy etc commenting is testament to that. So good stuff man.

I’m just putting the value for players ppm in excel and then putting that into a scatter plot. Then I just add a trend line and it tells me the equation.

Feedback is great and I’m glad that I’m getting some.
Thank you for the encouragement, hopefully I’m helping people (especially guys like mtb and variable kint at the top who are a chance of winning)

Ben Marlin wrote:
Thanks mate
How do you suggest that I could add depth and controls?

Sure, happy to help! I'll try and update this during the day, between meetings. This here is a helpful guide from a good source: https://eml.berkeley.edu/~dromer/papers/Econometrics%20Jan%202018.pdf

Ok, so this is going to be a very short discussion on regression.

You are interested understanding the relationship between a set of observations, and then predicting an outcome.

Each observation consists of a set of variables, which you seem to inherently understand well - i.e. Dude[z] played Minutes[x] against Team[y] in Position[p] in Round[r] with HIAtime[h] ScoringPoints[s]

Another way of saying this is that you have a model of what determines a player's scoring:
S = z + x + y + p + r + h

But that only works for one dude in one game, and it doesn't make any sense even then - because if we give positions a number between 1-17, then the model above says that reserves (14-17) will score higher. So we need to weight these things - in game t we have capital letter weights:

S(t) = D(z) + M(x) + T(y) + P(p) + R(r) + H(h)

but these guesses aren't perfectly correct - because sometimes a Dude will be tired, or really motivated, or have an undisclosed niggle, or play for Des Hasler. We have errors in the model:

S(t) = D(z) + M(x) + T(y) + P(p) + R(r) + H(h) + e(t)

So, if Cook scores 65 against the Dragons with no HIA, in hooker, for 80 minutes, in Round 19:

65 = D(Cookie, or player number 71 [we have to give a number]) + M(80) + T(14 [a number for Dragons]) + P(9) + R(18) + H(0) + e

The next trick is to figure out how big those capital letter weights need to be. We do this, often, by putting all the observations we have into a matrix, and minimising the squared errors in guesses: OLS.

Now, on to your question:
1) the more observations we have, the better our understanding of each Weighted letter will be
1a) this means that a calculation including all available games from Cook, Farah, Teddy, Wighton and BS9 will likely be more informative than just Cook - because we learn more about each Weighted letter with each observation.
2) We also need to think about how many explanatory poitns (eg H, P, M) we should use, as too many makes the calculation meaningless, and too few misses important stuff. The balancing line is a matter of art and science.

Happy to discuss more, but the Romer link is a good place to start. If you're really keen, there's a good uni near you, or lots of great free stuff online (MITx etc).

Cheers, and happy econometrics!

.....

That's great stuff L-Jimmy, especially for amateur guys like me. I'll have a proper read of the link later.

It's much easier to follow than I was expecting.

One of the things I've read about in predicting stuff like this, is not to use a thing to predict itself

ie, in Soccer, if you want to predict gals scored, rather than using goals scored in the past, they use a model that looks at things number of shots taken, and from what position those shots are taken, to come up with a better expectation of goals that will be scored than just looking at past history of what shots have actually got in.

Likewise for fantasy, using past points is not necessarily the best stats to use to predict future points.
Because its the most easily available stat, its what people tend to use, but even then people know enough to make adjustments
eg, DCE has lost goal kicking, take some points away
or makeup of the team may lead us to estimate more or less kick metres for a given player
or if someone has scored a lot of tries recently we might assume that the tries will dry up soon and adjust accordingly
We also look at factors like the opposition to make adjustments to predictions (eg Titans left hand defence is terrible, so right side attacking players might get a boost against them)

But basically the more factors we are able to consider, the more accurate our predictions can be.

Most of us don't have access to the complete stats required to make good predictions, or are skilled enough in using tools available to make predictions out of the stats we do have, so we make do with what we can. (But as L-Jimmy said, without using too many as to make the calculation meaningless - its about finding the factors that have the most influence on likely scores)

Milchy wrote:
One of the things I've read about in predicting stuff like this, is not to use a thing to predict itself

ie, in Soccer, if you want to predict gals scored, rather than using goals scored in the past, they use a model that looks at things number of shots taken, and from what position those shots are taken, to come up with a better expectation of goals that will be scored  than just looking at past history of what shots have actually got in.

Likewise for fantasy, using past points is not necessarily the best stats to use to predict future points.
Because its the most easily available stat, its what people tend to use, but even then people know enough to make adjustments
eg, DCE has lost goal kicking, take some points away
or makeup of the team may lead us to estimate more or less kick metres for a given player
or if someone has scored a lot of tries recently we might assume that the tries will dry up soon and adjust accordingly
We also look at factors like the opposition to make adjustments to predictions (eg Titans left hand defence is terrible, so right side attacking players might get a boost against them)

But basically the more factors we are able to consider, the more accurate our predictions can be.

Most of us don't have access to the complete stats required to make good predictions, or are skilled enough in using tools available to make predictions out of the stats we do have, so we make do with what we can. (But as L-Jimmy said, without using too many as to make the calculation meaningless - its about finding the factors that have the most influence on likely scores)

There's wisdom in using both, and I know where you're coming from.

Using a thing to predict itself relies on a phenomena called autocorrelation - and it can be pretty powerful in certain circumstances. You want to predict tomorrow's AUD/USD swap? Use today's (there'll be nothing much better!).

In specifying predictive models we nearly always use a little bit of autocorrelation, but not always at the first lag - for something you'll want to use 2, 3, 4, 6 or 12 lags (like with seasonal variations in snowfall, or mountain bike rental).

Where your insight into prediction is absolutely correct is where the autocorrelation is so powerful that it overwhelms other factors, and we become only interested in the rate of change (like GDP, or train delays). In these situations we 'difference' observations, and try and predict how different next week will be from this week (or how the averages and variances will change).

As for the 'access to complete stats ... tools available' bit - well, I'm in Sydney right now, but won't be travelling next week or the week after. @Milchy if you want to send me an xls with a question, I'm happy to run any regression you want - this includes using some automated tools to find the best model.

Same goes for the rest of you!

L-Jimmy I'll send you a PM later with where I am currently stuck with my data collection (It's more about importing from complex JSON into usable format than anything else)

Probably excel is not the best tool for what I want to do, but it's what I know.

Milchy wrote:
L-Jimmy I'll send you a PM later with where I am currently stuck with my data collection (It's more about importing from complex JSON into usable format than anything else)

Probably excel is not the best tool for what I want to do, but it's what I know.

Yeah, excel hits its statistical analysis limits pretty quickly - VBA is a bitch like that.

I use Stata, eViews and R. eViews has the lowest adoption curve, but unless you download it illegally you'll face a few hundred bucks in license fees.

I've done a couple of introductory R courses online, but not enough to do anything useful with it

Just a bit off topic. Can we take this to the nerd thread ?