Transformer Forecaster with neuralforecast¶

import matplotlib.pyplot as plt

import pandas as pd
import seaborn as sns

sns.set_theme(context="paper", style="ticks", palette="colorblind")

data_source = "https://gist.githubusercontent.com/emptymalei/15e548f483e896be0afc99e49dd6fbc9/raw/8384d82cdaaed4d9b0a555888efb70d79911dd2a/sunspot_nixtla.csv"

df = pd.read_csv(data_source, parse_dates=["ds"])

df.describe()

_, ax = plt.subplots()
sns.lineplot(df, x="ds", y="y", ax=ax)

ax.set_title("Sunspot Time Series")
ax.set_ylabel("Sunspot Number")

train_test_split_date = "2000-01-01"

df_train = df[df.ds <= train_test_split_date]
df_test = df[df.ds > train_test_split_date]

horizon = 3
horizon

from statsforecast import StatsForecast
from statsforecast.models import AutoARIMA

sf = StatsForecast(models=[AutoARIMA(season_length=12)], freq="YS")

sf.fit(df_train)

def forecast_test(forecaster_pred, df_train, df_test):
    dfs_pred = []
    for i in range(len(df_test)):
        df_pred_input_i = pd.concat([df_train, df_test[:i]])
        print(df_pred_input_i.ds.max())
        df_pred_output_i = forecaster_pred(df_pred_input_i)
        df_pred_output_i["step"] = i
        dfs_pred.append(df_pred_output_i)
    df_y_hat = pd.concat(dfs_pred).reset_index(drop=False)
    return df_y_hat

def visualize_predictions(
    df_train: pd.DataFrame,
    df_test: pd.DataFrame,
    df_pred: pd.DataFrame,
    model_name: str,
):
    _, ax = plt.subplots()

    sns.lineplot(df_train, x="ds", y="y", ax=ax)

    sns.lineplot(df_test, x="ds", y="y", color="k", ax=ax)

    sns.lineplot(df_pred, x="ds", y=model_name, hue="step", ax=ax)

    ax.set_title("Sunspot Forecast", fontsize=22)
    ax.set_ylabel("Sunspot Number", fontsize=20)
    ax.set_xlabel("Year", fontsize=20)
    ax.legend(prop={"size": 15})
    ax.grid()

df_y_hat_arima = forecast_test(
    lambda x: sf.predict(h=horizon, level=[95]), df_train=df_train, df_test=df_test
)

visualize_predictions(
    df_train=df_train, df_test=df_test, df_pred=df_y_hat_arima, model_name="AutoARIMA"
)

import matplotlib.pyplot as plt

fig, ax = plt.subplots(1, 1, figsize=(20, 7))
df_chart = df_test.merge(df_y_hat_arima, how="left", on=["unique_id", "ds"])
df_chart = pd.concat([df_train, df_chart]).set_index("ds")

df_chart[
    [
        "y",
        "AutoARIMA",
        # 'NHITS'
    ]
].plot(ax=ax, linewidth=2)

ax.set_title("AirPassengers Forecast", fontsize=22)
ax.set_ylabel("Monthly Passengers", fontsize=20)
ax.set_xlabel("Timestamp [t]", fontsize=20)
ax.legend(prop={"size": 15})
ax.grid()

from neuralforecast import NeuralForecast
from neuralforecast.models import NBEATS, NHITS, VanillaTransformer, iTransformer

models = [
    VanillaTransformer(
        # input_size=9, h=horizon,
        # n_head=4,
        # windows_batch_size=512,
        # learning_rate=0.00010614524276500768,
        # # conv_hidden_size=2,
        # # encoder_layers=6,
        # max_steps=500,
        # #early_stop_patience_steps=5,
        **{
            "hidden_size": 128,
            "n_head": 4,
            "learning_rate": 0.00010614524276500768,
            "scaler_type": "robust",
            "max_steps": 500,
            "batch_size": 32,
            "windows_batch_size": 512,
            "random_seed": 16,
            "input_size": 30,
            "step_size": 3,
            "h": horizon,
        }
    ),
    # iTransformer(
    #     input_size=history_length, h=horizon,
    #     # n_head=1,
    #     # conv_hidden_size=1,
    #     # encoder_layers=6,
    #     max_steps=50,
    #     n_series=1,
    #     #early_stop_patience_steps=5,
    # ),
    # NBEATS(input_size=history_length, h=horizon, max_steps=100, #early_stop_patience_steps=5
    #        ),
    # NHITS(input_size=history_length, h=horizon, max_steps=100, #early_stop_patience_steps=5
    #       )
]
nf = NeuralForecast(models=models, freq="YS")

nf.fit(df=df_train)

train_test_split_date

dfs_pred = []
for i in range(len(df_test)):
    df_pred_input_i = pd.concat([df_train, df_test[:i]])
    df_pred_output_i = nf.predict(df_pred_input_i)
    df_pred_output_i["step"] = i
    dfs_pred.append(df_pred_output_i)
df_y_hat = pd.concat(dfs_pred).reset_index(drop=False)

df_y_hat

df_test

df_y_hat.ds.dt.freq

import matplotlib.pyplot as plt

fig, ax = plt.subplots(1, 1, figsize=(20, 7))

sns.lineplot(df_train, x="ds", y="y", ax=ax)

sns.lineplot(df_test, x="ds", y="y", color="k", ax=ax)

sns.lineplot(df_y_hat, x="ds", y="VanillaTransformer", hue="step", ax=ax)


ax.set_title("Sunspot Forecast", fontsize=22)
ax.set_ylabel("Avg Sunspot Area", fontsize=20)
ax.set_xlabel("Year", fontsize=20)
ax.legend(prop={"size": 15})
ax.grid()

from datasetsforecast.evaluation import accuracy
from datasetsforecast.losses import mae, mse, rmse

df_y_hat.loc[df_y_hat.step == 0]

df_test

evaluation_df = accuracy(
    Y_hat_df=df_y_hat.loc[df_y_hat.step == 0],
    Y_test_df=df_test,
    Y_df=df_train,
    metrics=[mse, mae, rmse],
    agg_by=["unique_id"],
)

evaluation_df.head()