add EWM method for 3PAT averaging #1
Conversation
|
|
||
| fg3m_s = [] | ||
| #simulate n_simulations games | ||
| for _ in trange(n_simulated_games, desc=f'Simulating 3PM outcomes for {player_name} vs {opponent}...'): | ||
|
|
||
| #simulate FGA | ||
| fga_i = np.random.poisson(np.random.normal(fga_per_game_est_mean, fga_per_game_est_std)) | ||
| fga_i = np.random.poisson(max(np.random.normal(fga_per_game_est_mean, fga_per_game_est_std),0)) |
There was a problem hiding this comment.
So above you're not actually bootstrapping (for ema) so I don't think it's fair to draw from a normal distribution to input into the poisson. In fact I think this may be skewing the results for simulations quite a bit since the rolling std should be much larger than std of the bootstrapped emas. I could be wrong on that second part though.
| fga_per_game_est_std = np.std(fga_per_game_est) | ||
| elif fga_method == 'ewm': | ||
| fga_per_game_est_mean = fga_per_game_data['SHOT_ATTEMPTED_FLAG'].ewm(span = len(fga_per_game_data)).mean().values[-1] | ||
| fga_per_game_est_std = fga_per_game_data['SHOT_ATTEMPTED_FLAG'].ewm(span = len(fga_per_game_data)).std().values[-1] |
There was a problem hiding this comment.
No bootstrapping is getting done here, won't be able to compress into a list comprehension but I think we can write a for loop to do it:
eg.
bootstrapped_emas = []
for _ in range(bootstrap_samples):
fga_boot = np.random.choice(fga_per_game_array, size=len(fga_per_game_array), replace=True)
fga_boot_ema = pd.ewma(fga_boot, span=len(fga_boot)).mean()[-1]
bootstrapped_emas.append(fga_boot_ema)
fga_per_game_est_mean = np.mean(bootstrapped_means)
fga_per_game_est_std = np.std(bootstrapped_means)I've never actually done this before and I'd like to actually examine the distribution here in a notebook or something. Since EMA is dependent on the ordering I'm not sure this is valid, but I think it does work (mentioned this point earlier)
There was a problem hiding this comment.
Actually having messed around with this in a notebook i dont think it does work. ordering does matter
No description provided.