02. Busca de híbridos e seleção incremental¶

A busca roda em estágios, sempre na validação temporal:

Bases individuais de várias famílias (estatística, linear, kernel, árvore, neural).
Retenção: as duas melhores de cada família, o ARIMA e a melhor base geral.
Híbridos residuais: ŷ = base + α · corretor(resíduos), com corretores Linear, KNN e MLP, janelas de 7 ou 28 resíduos e α ∈ {0,5; 1}.
Refinamento das duas melhores estruturas residuais.
Seleção incremental: cada família desafia o campeão atual e só o substitui com ganho mediano ≥ 5% e vitórias em ≥ 60% das origens (regra calibrada, ADR-0008).

Perfil quickstart: o critério de variabilidade do protocolo fica desligado, e isso aparece no relatório. O perfil research_strict exige esse critério definido.

import autohybts

print("AutoHybTS", autohybts.__version__)

AutoHybTS 1.0.0rc4

1. Dados e experimento¶

O ARIMA com ordem automática é o modelo mais caro; aqui ele aparece só nos exemplos manuais da seção 3, e a busca automática fica com os demais modelos.

budget limita o número de candidatas (total, por família e por etapa; max_trials substitui o total), então a mesma série e a mesma semente dão a mesma busca em qualquer máquina. Um prazo de relógio é opcional (max_wall_time_s=...): nenhuma candidata começa depois dele, e as que ficarem de fora aparecem no ledger como STOPPED.

progress=True mostra uma barra do tqdm por etapa da busca e outra para as finalistas em V_select (extra progress). É só relatório: as candidatas avaliadas e o campeão não mudam.

from autohybts import AutoHybridExperiment
from autohybts.datasets import get_data

y = get_data("synthetic_daily")

exp = AutoHybridExperiment(
    fh=14,
    session_id=42,
    progress=True,
    fold=3,
    fold_strategy="expanding",
    seasonal_period=7,
    profile="quickstart",
    base_models=["naive", "seasonal_naive", "linear", "knn", "mlp"],
    correctors=["linear", "knn", "mlp"],
    residual_windows=[7, 28],
    residual_sources=["historical_oof_1step"],
    alphas=[0.5, 1.0],
    budget="small",
    max_trials=40,
).fit(y)
exp.splits

	layer	train_start	origin	first_target	last_target	n_train	horizon
fold_id
search-0	search	2024-01-01	2025-09-23	2025-09-24	2025-10-07	632	14
search-1	search	2024-01-01	2025-10-07	2025-10-08	2025-10-21	646	14
search-2	search	2024-01-01	2025-10-21	2025-10-22	2025-11-04	660	14
select-0	select	2024-01-01	2025-11-04	2025-11-05	2025-11-18	674	14
select-1	select	2024-01-01	2025-11-18	2025-11-19	2025-12-02	688	14
select-2	select	2024-01-01	2025-12-02	2025-12-03	2025-12-16	702	14

2. Catálogo de modelos¶

exp.models()

		family	multi_step	online_state_update	in_sample_fitted	deterministic	requires_extra	training_scope	output_type	horizon_strategy	max_horizon	target_mode	accepts_exog	requires	version	description
kind	name
base	adaboost_r2	tree	True	True	False	False	NaN	local	point	recursive	None	univariate	True	None	None	AdaBoost.R2 (scikit-learn) sobre lags.
	adida	intermittent	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	ADIDA: suavização em blocos de k períodos.
	arima	statistical	True	True	True	True	stats	local	point	recursive	None	univariate	True	None	None	ARIMA (statsmodels); ordem automática por KPSS...
	catboost_regressor	tree	True	True	False	False	boosting	local	point	recursive	None	univariate	True	None	None	CatBoost sobre lags (extra boosting).
	croston_sba	intermittent	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Croston-SBA: tamanho / intervalo entre demandas.
	elastic_net_regression	linear	True	True	False	True	NaN	local	point	recursive	None	univariate	True	None	None	Elastic Net sobre lags.
	exponential_smoothing	statistical	True	True	False	True	stats	local	point	recursive	None	univariate	False	None	None	Suavização exponencial aditiva (Holt, Holt-Win...
	extra_trees_regressor	tree	True	True	False	False	NaN	local	point	recursive	None	univariate	True	None	None	Árvores extremamente aleatórias sobre lags.
	harmonic_regression	statistical	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Regressão harmônica dinâmica: tendência + Four...
	hist_gradient_boosting	tree	True	True	False	False	NaN	local	point	recursive	None	univariate	True	None	None	HistGradientBoosting (scikit-learn) sobre lags.
	historic_average	statistical	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Média do treino (parâmetro congelado).
	imapa	intermittent	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	IMAPA: média do ADIDA nos níveis 1..k.
	knn	kernel	True	True	False	True	NaN	local	point	recursive	None	univariate	True	None	None	k vizinhos mais próximos sobre lags.
	lasso_regression	linear	True	True	False	True	NaN	local	point	recursive	None	univariate	True	None	None	Regressão Lasso sobre lags.
	lightgbm_regressor	tree	True	True	False	False	boosting	local	point	recursive	None	univariate	True	None	None	LightGBM sobre lags (extra boosting).
	linear	linear	True	True	False	True	NaN	local	point	recursive	None	univariate	True	None	None	Regressão linear sobre lags.
	mlp	neural	True	True	False	False	NaN	local	point	recursive	None	univariate	True	None	None	Perceptron multicamadas sobre lags.
	naive	statistical	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Repete o último valor observado.
	random_forest	tree	True	True	False	False	NaN	local	point	recursive	None	univariate	True	None	None	Floresta aleatória sobre lags.
	random_walk_drift	statistical	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Passeio aleatório com drift: último valor + h ...
	ridge_regression	linear	True	True	False	True	NaN	local	point	recursive	None	univariate	True	None	None	Regressão Ridge sobre lags.
	sarima	statistical	True	True	True	True	stats	local	point	recursive	None	univariate	True	None	None	ARIMA sazonal (statsmodels); D pela força sazo...
	seasonal_naive	statistical	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Repete a última temporada (season_length).
	seasonal_window_average	statistical	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Média de cada posição sazonal nas últimas wind...
	sf_auto_arima	statistical	True	True	False	True	nixtla-stats	local	point	recursive	None	univariate	False	None	None	AutoARIMA do StatsForecast (ordem automática, ...
	sf_auto_ces	statistical	True	True	False	True	nixtla-stats	local	point	recursive	None	univariate	False	None	None	AutoCES do StatsForecast (suavização exponenci...
	sf_auto_ets	statistical	True	True	False	True	nixtla-stats	local	point	recursive	None	univariate	False	None	None	AutoETS do StatsForecast (erro, tendência e sa...
	sf_auto_theta	statistical	True	True	False	True	nixtla-stats	local	point	recursive	None	univariate	False	None	None	AutoTheta do StatsForecast (variantes do métod...
	svm	kernel	True	True	False	True	NaN	local	point	recursive	None	univariate	True	None	None	SVR sobre lags.
	tree	tree	True	True	False	True	NaN	local	point	recursive	None	univariate	True	None	None	Árvore de decisão sobre lags.
	tsb	intermittent	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	TSB: probabilidade de demanda × tamanho.
	window_average	statistical	True	True	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Média das últimas window observações.
	xgboost_regressor	tree	True	True	False	False	boosting	local	point	recursive	None	univariate	True	None	None	XGBoost sobre lags (extra boosting).
corrector	adaboost_r2	corrector	False	False	False	False	NaN	local	point	recursive	None	univariate	False	None	None	AdaBoost.R2 sobre janelas de resíduos.
	arima_sequence	corrector	True	True	False	True	stats	local	point	recursive	None	univariate	False	None	None	ARIMA sobre a série de resíduos (horizonte int...
	catboost_regressor	corrector	False	False	False	False	boosting	local	point	recursive	None	univariate	False	None	None	CatBoost sobre janelas de resíduos (extra boos...
	exponential_smoothing_sequence	corrector	True	True	False	True	stats	local	point	recursive	None	univariate	False	None	None	Suavização exponencial sobre a série de resídu...
	hist_gradient_boosting	corrector	False	False	False	False	NaN	local	point	recursive	None	univariate	False	None	None	HistGradientBoosting sobre janelas de resíduos.
	knn	corrector	False	False	False	True	NaN	local	point	recursive	None	univariate	False	None	None	k vizinhos mais próximos sobre janelas de resí...
	lightgbm_regressor	corrector	False	False	False	False	boosting	local	point	recursive	None	univariate	False	None	None	LightGBM sobre janelas de resíduos (extra boos...
	linear	corrector	False	False	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Regressão linear sobre janelas de resíduos.
	mlp	corrector	False	False	False	False	NaN	local	point	recursive	None	univariate	False	None	None	Perceptron multicamadas sobre janelas de resíd...
	random_forest	corrector	False	False	False	False	NaN	local	point	recursive	None	univariate	False	None	None	Floresta aleatória sobre janelas de resíduos.
	svm	corrector	False	False	False	True	NaN	local	point	recursive	None	univariate	False	None	None	SVR sobre janelas de resíduos.
	tree	corrector	False	False	False	True	NaN	local	point	recursive	None	univariate	False	None	None	Árvore de decisão sobre janelas de resíduos.
	xgboost_regressor	corrector	False	False	False	False	boosting	local	point	recursive	None	univariate	False	None	None	XGBoost sobre janelas de resíduos (extra boost...

3. Um modelo individual e um híbrido montado à mão¶

Útil para entender as peças antes da busca automática. Os resíduos do corretor são erros de previsão fora da amostra (historical_oof_1step), nunca valores ajustados no próprio treino.

arima = exp.create_model("arima")
arima.cv_results

	layer	status	MASE	MAE	RMSE	sMAPE
search-0	search	SUCCEEDED	1.846794	2.865174	3.100497	0.040837
search-1	search	SUCCEEDED	0.554999	0.861073	1.100535	0.012383
search-2	search	SUCCEEDED	1.668095	2.569822	2.698243	0.037090
select-0	select	SUCCEEDED	2.174782	3.346846	3.488244	0.047806
select-1	select	SUCCEEDED	2.195924	3.418777	3.807198	0.048332
select-2	select	SUCCEEDED	0.387069	0.607182	0.739234	0.008558
Mean	NaN	NaN	1.471277	2.278146	2.488992	0.032501
Std	NaN	NaN	0.801855	1.238794	1.276195	0.017626

hybrid = exp.create_hybrid(
    base="arima",
    corrector="mlp",
    window=7,
    alpha=1.0,
    residual_source="historical_oof_1step",
)
hybrid.cv_results

	layer	status	MASE	MAE	RMSE	sMAPE
search-0	search	SUCCEEDED	1.866445	2.895661	3.093652	0.041158
search-1	search	SUCCEEDED	0.555764	0.862260	1.086767	0.012409
search-2	search	SUCCEEDED	1.668606	2.570609	2.692855	0.037114
select-0	select	SUCCEEDED	2.083504	3.206376	3.343848	0.045730
select-1	select	SUCCEEDED	2.161925	3.365846	3.776502	0.047617
select-2	select	SUCCEEDED	0.365805	0.573826	0.722414	0.008059
Mean	NaN	NaN	1.450342	2.245763	2.452673	0.032015
Std	NaN	NaN	0.787974	1.217761	1.254970	0.017317

4. Busca automática¶

result = exp.compare_models(include=["standalone", "residual"], sort="MASE")
result.leaderboard

busca S0 bases:   0%|          | 0/5 [00:00<?, ?cand/s]

busca S0 bases:  80%|████████  | 4/5 [00:00<00:00, 25.08cand/s]

busca S0 bases: 100%|██████████| 5/5 [00:00<00:00,  5.07cand/s]

busca S2 resíduos:   0%|          | 0/11 [00:00<?, ?cand/s]

busca S2 resíduos:   9%|▉         | 1/11 [00:02<00:20,  2.08s/cand]

busca S2 resíduos:  18%|█▊        | 2/11 [00:03<00:13,  1.52s/cand]

busca S2 resíduos:  27%|██▋       | 3/11 [00:03<00:07,  1.11cand/s]

busca S2 resíduos:  36%|███▋      | 4/11 [00:03<00:05,  1.35cand/s]

busca S2 resíduos:  45%|████▌     | 5/11 [00:04<00:03,  1.85cand/s]

busca S2 resíduos:  55%|█████▍    | 6/11 [00:05<00:04,  1.21cand/s]

busca S2 resíduos:  64%|██████▎   | 7/11 [00:06<00:03,  1.15cand/s]

busca S2 resíduos:  73%|███████▎  | 8/11 [00:06<00:01,  1.54cand/s]

busca S2 resíduos:  82%|████████▏ | 9/11 [00:07<00:01,  1.64cand/s]

busca S2 resíduos:  91%|█████████ | 10/11 [00:07<00:00,  2.04cand/s]

busca S2 resíduos: 100%|██████████| 11/11 [00:08<00:00,  1.23cand/s]

busca S2 resíduos: 100%|██████████| 11/11 [00:08<00:00,  1.24cand/s]

busca S3 refino:   0%|          | 0/2 [00:00<?, ?cand/s]

busca S3 refino:  50%|█████     | 1/2 [00:02<00:02,  2.08s/cand]

busca S3 refino: 100%|██████████| 2/2 [00:02<00:00,  1.05cand/s]

busca S3 refino: 100%|██████████| 2/2 [00:02<00:00,  1.12s/cand]

seleção V_select:   0%|          | 0/5 [00:00<?, ?cand/s]

seleção V_select:  20%|██        | 1/5 [00:00<00:01,  2.12cand/s]

seleção V_select:  40%|████      | 2/5 [00:01<00:02,  1.07cand/s]

seleção V_select: 100%|██████████| 5/5 [00:01<00:00,  2.80cand/s]

	name	model	family	strategy	base	corrector	residual_source	window	alpha	mean_rank	...	MAE	RMSE	sMAPE	coverage	n_units	n_failed	n_skipped	fit_seconds	predict_seconds	is_champion
candidate_id
7dd1082a4ee4fe9df20683376c5688b4a23322240929c4c44a8140e266a10ae2	mlp	mlp(lags=7)	neural	standalone	mlp	NaN	NaN	NaN	NaN	1.0	...	1.329098	1.590181	0.019224	1.0	3	0	0	0.807899	0.017187	False
9fd962cca51322c0e3697027ac96d1fa8d35c2672647b0030737ef663dde6df6	mlp-linear	mlp(lags=7) + linear [w=7, alpha=0.5, historic...	residual	residual_additive	mlp	linear	historical_oof_1step	7.0	0.5	2.0	...	1.335050	1.622713	0.019335	1.0	3	0	0	2.061273	0.012969	True
459af2c77fe43d1936669c337e695b972b6de79264cfa329ed349dfc20dc5172	mlp-linear	mlp(lags=7) + linear [w=28, alpha=0.5, histori...	residual	residual_additive	mlp	linear	historical_oof_1step	28.0	0.5	3.0	...	1.536977	1.800618	0.022398	1.0	3	0	0	1.533249	0.012512	False
70ad174e1cc49eb256646399fb3e4b7d9df82f14f2978d09d35dac0019c7ef37	knn	knn(lags=7)	kernel	standalone	knn	NaN	NaN	NaN	NaN	4.0	...	1.563061	1.966950	0.022621	1.0	3	0	0	0.050587	0.036638	False
d90528888825a4459816afbcf249ff909747924f0a2bbaa6e09d027e31342238	seasonal_naive	seasonal_naive(season_length=7)	statistical	standalone	seasonal_naive	NaN	NaN	NaN	NaN	5.0	...	1.614291	1.909265	0.023346	1.0	3	0	0	0.000080	0.000144	False
e9e747a9648ad8d377f20122dce611a43dd3bf85a64c950a8e1ef246ebc032b4	linear	linear(lags=7)	linear	standalone	linear	NaN	NaN	NaN	NaN	6.0	...	1.649719	1.852607	0.023763	1.0	3	0	0	0.057596	0.008549	False
773d3b8c8d625571d4d4b4f4aa234d9f88b9b7fc0ceb3101c126e49e100378fc	seasonal_naive-linear	seasonal_naive(season_length=7) + linear [w=7,...	residual	residual_additive	seasonal_naive	linear	historical_oof_1step	7.0	0.5	7.0	...	1.673298	1.998234	0.024205	1.0	3	0	0	0.164522	0.005550	False
c4821c0a39e39a26ca17be31d8298f1ebcaed90d98e0a1fd3722f7c6ed592f9b	linear-linear	linear(lags=7) + linear [w=7, alpha=0.5, histo...	residual	residual_additive	linear	linear	historical_oof_1step	7.0	0.5	8.0	...	1.698826	1.914754	0.024472	1.0	3	0	0	0.481665	0.012821	False
972a9d57c62d5265bf431a410bf8ccd45ae226f2e6bb124bc9d763258cbdbae4	mlp-linear	mlp(lags=7) + linear [w=7, alpha=1.0, historic...	residual	residual_additive	mlp	linear	historical_oof_1step	7.0	1.0	9.0	...	1.725096	2.009527	0.024833	1.0	3	0	0	1.366397	0.012967	False
b678627c69f0bc5eb419617a2119103a8a5340ac96c0551898c62a1d76dc0179	linear-linear	linear(lags=7) + linear [w=7, alpha=1.0, histo...	residual	residual_additive	linear	linear	historical_oof_1step	7.0	1.0	10.0	...	1.748087	1.988548	0.025178	1.0	3	0	0	0.504137	0.018069	False
507265ea41efe47e82119ba51bc162c1cd983e9fd72f0a1b25e7ca1c8d95b229	seasonal_naive-linear	seasonal_naive(season_length=7) + linear [w=7,...	residual	residual_additive	seasonal_naive	linear	historical_oof_1step	7.0	1.0	11.0	...	1.768259	2.152438	0.025573	1.0	3	0	0	0.159355	0.005538	False
bbd8aa955229e78f79ee95ec40b350010922f578702fd6a464ea2ea729023f34	seasonal_naive-linear	seasonal_naive(season_length=7) + linear [w=7,...	residual	residual_additive	seasonal_naive	linear	historical_oof_multistep	7.0	0.5	12.0	...	2.617302	2.887196	0.037349	1.0	3	0	0	0.148647	0.006292	False
60fa40d778d88419b30bab18cd7f1fd7cad8e46e100cfe53d3a7b3f2fcc845b4	naive-linear	naive + linear [w=7, alpha=1.0, historical_oof...	residual	residual_additive	naive	linear	historical_oof_1step	7.0	1.0	13.0	...	3.903393	4.441350	0.055259	1.0	3	0	0	0.210182	0.009204	False
b78538dee2cc85795a8cc90439de94a964e7a02a28eff7a3f960c947a2e9e699	naive-linear	naive + linear [w=7, alpha=0.5, historical_oof...	residual	residual_additive	naive	linear	historical_oof_1step	7.0	0.5	14.0	...	3.916747	4.375241	0.055684	1.0	3	0	0	0.174942	0.005655	False
2117d1027abc8c33b855cff511914a6917df041e30dd98281ad26a716bae3b78	naive	naive	statistical	standalone	naive	NaN	NaN	NaN	NaN	15.0	...	3.961290	4.498781	0.056375	1.0	3	0	0	0.000119	0.000137	False
bf5bff5cb0bc3a6ce953ea9188dae5a44f9f780d655ab31857957fa89a1f2250	knn-linear	knn(lags=7) + linear [w=7, alpha=0.5, historic...	residual	residual_additive	knn	linear	historical_oof_1step	7.0	0.5	16.0	...	4.072110	4.339678	0.057173	1.0	3	0	0	1.077592	0.042116	False
747968575a6cd0d95bbee95e83321806fed5a85e1007f47989c2adf289a7f495	mlp-linear	mlp(lags=7) + linear [w=7, alpha=0.5, historic...	residual	residual_additive	mlp	linear	historical_oof_multistep	7.0	0.5	17.0	...	4.429702	4.649585	0.061832	1.0	3	0	0	2.063942	0.014461	False
df3ec82ac4a2623a390d9a7ee621e53fda49f4f8b54f4d411b38c569f7b2af05	knn-linear	knn(lags=7) + linear [w=7, alpha=1.0, historic...	residual	residual_additive	knn	linear	historical_oof_1step	7.0	1.0	18.0	...	8.484694	8.671081	0.115437	1.0	3	0	0	0.925168	0.036879	False

18 rows × 21 columns

5. Por que o campeão venceu¶

decisions lista cada desafiante, se foi aceito, o ganho mediano, a taxa de vitórias e os motivos. Rejeições também ficam registradas. audit conta todas as tentativas, inclusive as podadas e as que falharam.

Este notebook usa fold=3 para rodar rápido: são 3 unidades em V_select, e 60% de vitórias exige vencer 2 das 3. Com tão poucas unidades a regra é um protocolo, não um teste estatístico. O padrão fold="auto" usa 7 origens quando a série comporta, o número em que a regra foi calibrada; o notebook 11 mede quantas vezes cada configuração aceita um híbrido espúrio.

result.decisions

	family	champion_before	challenger	accepted	reason_codes	median_gain	win_rate	variability_status	variability_value	p_value	challenger_coverage	challenger_label	champion_before_id	challenger_id
0	statistical	mlp	seasonal_naive	False	(median_gain_below_threshold,)	0.034103	0.666667	disabled	None	None	1.0	seasonal_naive(season_length=7)	7dd1082a4ee4fe9df20683376c5688b4a23322240929c4...	d90528888825a4459816afbcf249ff909747924f0a2bba...
1	linear	mlp	linear	False	(median_gain_below_threshold, win_rate_below_t...	-0.026810	0.333333	disabled	None	None	1.0	linear(lags=7)	7dd1082a4ee4fe9df20683376c5688b4a23322240929c4...	e9e747a9648ad8d377f20122dce611a43dd3bf85a64c95...
2	kernel	mlp	knn	False	(median_gain_below_threshold, win_rate_below_t...	-0.109042	0.333333	disabled	None	None	1.0	knn(lags=7)	7dd1082a4ee4fe9df20683376c5688b4a23322240929c4...	70ad174e1cc49eb256646399fb3e4b7d9df82f14f2978d...
3	residual	mlp	mlp-linear	True	(accepted,)	0.142037	0.666667	disabled	None	None	1.0	mlp(lags=7) + linear [w=7, alpha=0.5, historic...	7dd1082a4ee4fe9df20683376c5688b4a23322240929c4...	9fd962cca51322c0e3697027ac96d1fa8d35c2672647b0...

Tudo o que cada decisão usou, unidade por unidade: os dois erros, o ganho e se houve vitória (empate não conta). Dá para recalcular o ganho mediano e a taxa de vitórias, ou reamostrar, fora da biblioteca.

units = result.selection.decision_units_frame()
units[units["family"] == "residual"].round(4)

	family	challenger_id	accepted	unit_id	series_id	champion_error	challenger_error	gain	win
9	residual	9fd962cca51322c0e3697027ac96d1fa8d35c2672647b0...	True	select-0	None	2.3449	2.0119	0.1420	True
10	residual	9fd962cca51322c0e3697027ac96d1fa8d35c2672647b0...	True	select-1	None	1.2018	1.6898	-0.4061	False
11	residual	9fd962cca51322c0e3697027ac96d1fa8d35c2672647b0...	True	select-2	None	1.5715	1.1243	0.2846	True

result.audit

SelectionAudit(protocol_hash='aa88c5a5ada48e7970d42ca6d46bace46c42668c9f6b3ad47986044ed0787f56', search_protocol_hash='dd40456eb2da818cf575c202bb45b3c9f8473119d8df55ac0ae006dbd977e0d3', total_trials=18, tried_by_family={'statistical': 2, 'linear': 1, 'kernel': 1, 'neural': 1, 'residual': 13}, pruned_trials=49, failed_trials=0, finalist_ids=('7dd1082a4ee4fe9df20683376c5688b4a23322240929c4c44a8140e266a10ae2', '9fd962cca51322c0e3697027ac96d1fa8d35c2672647b0030737ef663dde6df6', '70ad174e1cc49eb256646399fb3e4b7d9df82f14f2978d09d35dac0019c7ef37', 'd90528888825a4459816afbcf249ff909747924f0a2bbaa6e09d027e31342238', 'e9e747a9648ad8d377f20122dce611a43dd3bf85a64c950a8e1ef246ebc032b4'), winner_id='9fd962cca51322c0e3697027ac96d1fa8d35c2672647b0030737ef663dde6df6', comparison_origin_ids=('select-0', 'select-1', 'select-2'), decision_reasons=('statistical: median_gain_below_threshold', 'linear: median_gain_below_threshold+win_rate_below_threshold', 'kernel: median_gain_below_threshold+win_rate_below_threshold', 'residual: accepted'), exploratory=False, search_best_score=0.8585660174722802, winner_search_score=0.8628987716271869, winner_select_score=1.6086533406757795)

6. Diagnósticos da validação¶

Todos os gráficos usam previsões fora da amostra nas origens de validação, nunca o teste selado.

exp.plot_model(result.best, plot="cv_forecasts")

../../_images/978d69942343de07531223e4b76e2e710e91225a9831ef414fc4b6bc7ea7e0b4.png

exp.plot_model(result.best, plot="residuals")

../../_images/670f2106fea4d6d67de79672993f4a5c9d5232fbd58b18a78b9943c773e910ee.png

exp.plot_model(result.best, plot="selection_trace")

../../_images/a85103568cc7375a57af368194f31fa0a3e7ff0a929e354bcc36a5b617b54733.png

7. Eventos do experimento¶

for event in exp.events:
    duration = f"{event.duration_ms:9.1f} ms" if event.duration_ms else " " * 12
    print(f"{event.kind:32s} {duration}  {event.message}")

experiment.fitted                     38.7 ms  1 série(s), 716 pontos de desenvolvimento, teste selado de 14 por série, 6 origens
model.created                      30679.7 ms  arima
hybrid.created                     21592.0 ms  arima + mlp [w=7, alpha=1.0, historical_oof_1step]
search.finished                    12167.0 ms  18 tentativas (budget_limited)
selection.finished                  1817.0 ms  campeão: mlp(lags=7) + linear [w=7, alpha=0.5, historical_oof_1step]