RandomForestRegressor predicts constant values

[tavlox]

I am trying to fit RandomForestRegressor on the training dataset, when using the predict function to predict the values, I receive constant values through time. What is the reason for this? My goal is to forecast up to 72 hours ahead.
Here is my code:
`from data_preparation import Preparation
from missing_timestamps import remove_duplicates
import pandas as pd
from skforecast.ForecasterAutoreg import ForecasterAutoreg
from skforecast.model_selection import grid_search_forecaster
from sklearn.ensemble import RandomForestRegressor
from skforecast.utils import save_forecaster
from skforecast.utils import load_forecaster

marvin

data = Preparation(r'/home/ieftimska/operato-meteo-1/data/MAS_processed/ELES-MAS-5001.csv.gz', "AMBIENT_TEMPERATURE")
#data = Preparation(r'/home/iva/Desktop/operato-meteo-1/data/MAS_processed/ELES-MAS-5001.csv.gz', "AMBIENT_TEMPERATURE")
train, test = data.split()
train_processed = remove_duplicates(train)
#train_processed_ = train_processed["AMBIENT_TEMPERATURE"].copy().squeeze()
test_processed = remove_duplicates(test)
#test_processed_ = test_processed["AMBIENT_TEMPERATURE"].copy().squeeze()
whole_data = pd.concat([train_processed, test_processed])
whole_data = whole_data.rename(columns={"AMBIENT_TEMPERATURE": "y"})
whole_data.index = whole_data.index.rename("datetime")
forecaster.fit(y=whole_data.loc[:"2022", "y"])
save_forecaster(forecaster, file_name='forecaster_random_forest.py', verbose=False)
forecaster_loaded = load_forecaster('forecaster_random_forest.py', verbose=True)
predictions = forecaster_loaded.predict(steps=864)Here is the dataset: [ELES-MAS-5001.csv.gz](https://github.com/JoaquinAmatRodrigo/skforecast/files/12646861/ELES-MAS-5001.csv.gz) This is the missing_timestamps script:import pandas as pd
from datetime import timedelta
import numpy as np

def remove_duplicates(data):
"""
A function that removes duplicates in timestamps and removes
timezone information from the timestamps
"""
data["timestamp"] = pd.to_datetime(data["timestamp"])
time = data["timestamp"].dt.tz_localize(None)
data["timestamp"] = time
data_processed = data.drop_duplicates(subset="timestamp", keep='last')
data_processed = data_processed.set_index("timestamp")
data_processed = data_processed.sort_index()

return data_processed

def missing_data(data_processed):
"""
A function that finds the dates of the missing data and fills it with the previous non-missing timestamp
for a better visualization to present train/test data with the missing data
"""
timestamps_series = pd.Series(data_processed.index)
diff_5 = timedelta(hours=0, minutes=5)
diff = timestamps_series.diff()
more_than_5 = np.where(diff > diff_5)[0]
previous_timestamp = more_than_5 - 1
gaps = list()
# a list of lists, where the first element is the previous timestamp of the missing data (succesive timestamps differ more than 5 min)
# and the second element is the end date of the missing data
for i, j in zip(data_processed.iloc[previous_timestamp].index, data_processed.iloc[more_than_5].index):
gaps.append([i, j])
missing_timestamps = list()
# dates of missing data where the start date is not a missing data
for i in gaps:
missing_timestamps.append(pd.date_range(start=i[0], end=i[1], freq='5T'))
missing_timestamps_one_by_one = list()
for i in missing_timestamps:
# exclude the non-missing timestamp
for j in i[1:]:
missing_timestamps_one_by_one.append(j)
values = data_processed.iloc[previous_timestamp].values
values_list = list()
for i, v in enumerate(values):
values_list.append(v[0])
# fill the missing timestamps with the values of their previous non-missing timestamps
missing_values_filled_previous = list()
for t, v in zip(missing_timestamps, values_list):
missing_values_filled_previous.append([v] * len(t[1:]))
missing_values_filled_previous_one_by_one = list()
for i in missing_values_filled_previous:
for j in i:
missing_values_filled_previous_one_by_one.append(j)

return missing_timestamps_one_by_one, missing_values_filled_previous_one_by_one`

[JoaquinAmatRodrigo]

We are closing this issue due to inactivity. If you believe this issue is still relevant, please feel free to reopen it or create a new issue with updated details.

Thank you for your understanding and cooperation.