MathiasH

Data Scientist. Specializing in development of statistical and machine-learning models intended for sports-betting.

[MathiasH @Hiderdk]

@craptor How do you measure CLV when you backtest? It requires that your model only has information that's available at the time of the bet, i.e. injury or other lineup news that may first be available at kickoff.

[MathiasH @Hiderdk]

@PlusEVAnalytics @d_feustel If you can sgp it its most likely done through a sim. The variance around the players skill might not be taking into account though.

[MathiasH @Hiderdk]

@PlusEVAnalytics @d_feustel What makes you think this isnt already priced by the sportsbook model?

[MathiasH @Hiderdk]

@SethWalder It also depends on your prediction for what Texans will do with the ball if Chiefs punt it. What are you predicting there?

[MathiasH @Hiderdk]

@PFF_Moo It's not a sim-based model? Are there any public-live sim-based NFL models?

[MathiasH @Hiderdk]

@JackJGaming In theory this could be solved by a a live-win probability model. Even something like tempo could be measured through positional coordinates (although complex).
The bigger problem is that it's harder to attribute (skill) contribution to an action.

[MathiasH @Hiderdk]

@ChristophMolnar It increases maintenance costs and effectively increases the length of your functions. All else being equal, less is better than more. There can be reasons to add comments, however, by default it's a code smell.

[MathiasH @Hiderdk]

@thepowerrank All models have biases. And to address these biases you need to add additional parameters to the model. It's in my opinion the most complex part of rating models to fully solve.

[MathiasH @Hiderdk]

@thepowerrank On the topic of whether rating models take into account skill differences between North American regions and european countries (soccer national teams). You replied it does it through pure math. However, from my experience, weak isolated regions tend to have inflated ratings.

[MathiasH @Hiderdk]

@PlusEVAnalytics I think if you add the handcrafted regresssion predictions as a feature to the xgboost (along with more raw features) that might be better. The reason is that it allows xgboost to identify in which cases regression might be better and vice versa.

[MathiasH @Hiderdk]

@marktenenholtz transformers can work for time-series without that much data. Temporal Fusion Transformer can generate close to comparable performance to gradient boosting.

[MathiasH @Hiderdk]

@RufusPeabody Data Scientists working on sports prediction models are being held to a significant higher standard than 99% of data scientists in other industries. In other industries domain knowledge is much lower and people can't identify the obvious edge-case errors.

[MathiasH @Hiderdk]

@spanky Someone might simply be risk averse or having less initial bank roll. Doesn't make them less sharp. While beating $300 limit by itself doesn't make you sharp. If you do it for +15% ROI, that's certainly impressive and quite different from being it by 3%.

[MathiasH @Hiderdk]

@mikaelmilhoj @tLindegaardN Men pengene bliver jo forrentet og derved tages der højde for tidsværdien af pengene. Derimod kan man sige at lønmodtagerne ikke får muligheden for selv at bestemme hvordan pengene skal bruges, hvilket har en vis form for værdi.

[MathiasH @Hiderdk]

@marktenenholtz You end up with too high learning rate when you tune it based on default n estimators. You get better performance with lower learning rate and tuning n estimators based on that

[MathiasH @Hiderdk]

@ASchatzNFL @IAmDaltos How do you measure improvements?

[MathiasH @Hiderdk]

@JackJGaming It's not perfect but reasonably accurate from my experience and much more time-efficient. For instance optimal max_depth is unlikely to be at 10 at 0.01 and 2 at 0.1. Hyperparameter tuning everything at 0.01 learning_rate is very slow.

[MathiasH @Hiderdk]

@JackJGaming A more efficient approach is to optimise all other hyperparameters using a high fixed learning_rate (e.g. 0.1). That will lead to fewer trees being optimal and thus your model will train much faster. In your final optimisation you tune n_estimators using a 0.01 learning_rate.

[MathiasH @Hiderdk]

@NateSilver538 @lxeagle17 I suggest you look at esports as case study. US had plenty of money due to vc capital and thus higher salaries but could not compete with regions where the the games were more popular. Dont see mls becoming a top league unless soccer becomes significantly more popular.