[Python] ARIMA Model give 0s for prediction when differentiating non Stationarity data

I am trying to learn ARIMA and so far I am working on this dataset

Huge Stock Market Dataset

i used intel stocks as you can see from the code, I was able to make it work with LSTM and that went smooth, but i dont have any experience with ARIMA

my current problem is the forecast is a flat line, and can't figure out the problem


import numpy as np
import pandas as pd
from pmdarima import auto_arima
from statsmodels.tsa.stattools import adfuller
from statsmodels.tsa.arima.model import ARIMA
from sklearn.metrics import mean_squared_error, mean_absolute_error
import matplotlib.pyplot as plt
import warnings

warnings.filterwarnings("ignore")

# Load and preprocess data
def load_data(filepath, start_date):
df = pd.read_csv(filepath, sep=',')
df = df[df['Date'] > start_date]
df['Date'] = pd.to_datetime(df['Date'])
df.set_index('Date', inplace=True)
return df[['Close']]

# ADF Test
def adf_test(series):
result = adfuller(series)
print(f"ADF Statistic: {result[0]}, p-value: {result[1]}")
return result[1] <= 0.05 # Stationary if p-value <= 0.05

# Plot data
def plot_series(series, title):
plt.figure(figsize=(12, 6))
plt.plot(series, label=title, color='blue', alpha=0.7)
plt.title(title)
plt.legend()
plt.show()

# Split data
def split_data(data, split_ratio=0.8):
train_size = int(len(data) * split_ratio)
train, test = data[:train_size], data[train_size:]
return train, test

# Fit ARIMA model
def fit_arima(train, order=None):
if not order:
stepwise_fit = auto_arima(train, seasonal=False, trace=True, error_action='ignore', suppress_warnings=True)
order = stepwise_fit.order
print(f"Selected ARIMA Order: {order}")
model = ARIMA(train, order=order)
return model.fit()

# Forecast and evaluate
def forecast_and_evaluate(model_fit, test, original_series=None, diff_order=0):
forecast = model_fit.forecast(steps=len(test))
forecast = pd.Series(forecast, index=test.index, name='Forecast')

# Reverse differencing if applicable
if diff_order > 0 and original_series is not None:
last_train_value = original_series.iloc[-1] # Last value from original (non-differenced) training series
forecast = forecast.cumsum() + last_train_value

# Debugging forecast values
print(f"First 5 forecast values (before adjustments):\n{forecast.head()}")

# Remove negative values
forecast[forecast < 0] = 0

# Remove non-numeric values
forecast = forecast.fillna(0)

# Calculate metrics
rmse = np.sqrt(mean_squared_error(test, forecast))
mae = mean_absolute_error(test, forecast)
print(f"\nRoot Mean Squared Error: {rmse}")
print(f"Mean Absolute Error: {mae}")

return forecast, rmse, mae

# Plot comparison
def plot_forecast(test, forecast, title='Actual vs Forecast'):
plt.figure(figsize=(12, 6))
plt.plot(test, label='Actual Test Data', color='blue', alpha=0.7)
plt.plot(forecast, label='Forecast', color='red', linestyle='--', alpha=0.7)
plt.title(title)
plt.legend()
plt.show()

# Residual diagnostics
def residual_diagnostics(residuals):
plt.figure(figsize=(12, 6))
plt.subplot(211)
plt.plot(residuals, label='Residuals', color='purple', alpha=0.7)
plt.title('Residuals')
plt.legend()
plt.subplot(212)
plt.hist(residuals, bins=20, alpha=0.7, color='green', label='Residual Histogram')
plt.title('Residual Histogram')
plt.legend()
plt.tight_layout()
plt.show()

# Main script
if __name__ == "__main__":
# Load data
filepath = "../Datasets/Stocks/intc.us.txt"
data = load_data(filepath, start_date='1990-01-01')

# Check stationarity
if not adf_test(data['Close']):
print("Data is non-stationary. Differencing...")
data_diff = data['Close'].diff().dropna()
diff_order = 1
print(f"First 5 values after differencing:\n{data_diff.head()}")
else:
print("Data is stationary.")
data_diff = data['Close']
diff_order = 0

plot_series(data_diff, "Differenced Data")

# Split data
train, test = split_data(data_diff)
original_train, original_test = split_data(data['Close'])

# Fit ARIMA model
model_fit = fit_arima(train)

# Residual diagnostics
residual_diagnostics(model_fit.resid)

# Forecast and evaluate
forecast, rmse, mae = forecast_and_evaluate(
model_fit,
test,
original_series=original_train if diff_order > 0 else None,
diff_order=diff_order
)

# Plot forecast vs actual
plot_forecast(test, forecast, title="Actual vs Forecast")

# Print forecast head
print(forecast.head())

here is the log

ADF Statistic: -1.2717375543756217, p-value: 0.642001214544929
Data is non-stationary. Differencing...
First 5 values after differencing:
Date
1990-01-03 -0.0242
1990-01-04 0.0164
1990-01-05 -0.0084
1990-01-08 0.0162
1990-01-09 0.0240
Name: Close, dtype: float64
Performing stepwise search to minimize aic
ARIMA(2,0,2)(0,0,0)[0] : AIC=9193.905, Time=0.56 sec
ARIMA(0,0,0)(0,0,0)[0] : AIC=9200.647, Time=0.08 sec
ARIMA(1,0,0)(0,0,0)[0] : AIC=9198.553, Time=0.06 sec
ARIMA(0,0,1)(0,0,0)[0] : AIC=9198.409, Time=0.08 sec
ARIMA(1,0,2)(0,0,0)[0] : AIC=9192.284, Time=0.35 sec
ARIMA(0,0,2)(0,0,0)[0] : AIC=9198.467, Time=0.12 sec
ARIMA(1,0,1)(0,0,0)[0] : AIC=9191.738, Time=0.23 sec
ARIMA(2,0,1)(0,0,0)[0] : AIC=9192.008, Time=0.29 sec
ARIMA(2,0,0)(0,0,0)[0] : AIC=9199.273, Time=0.09 sec
ARIMA(1,0,1)(0,0,0)[0] intercept : AIC=9193.371, Time=0.64 sec

Best model: ARIMA(1,0,1)(0,0,0)[0]
Total fit time: 2.495 seconds
Selected ARIMA Order: (1, 0, 1)
First 5 forecast values (before adjustments):
Date
2012-04-16 NaN
2012-04-17 NaN
2012-04-18 NaN
2012-04-19 NaN
2012-04-20 NaN
Name: Forecast, dtype: float64

Root Mean Squared Error: 0.38750459255687414
Mean Absolute Error: 0.2705679715302491
Date
2012-04-16 0.0
2012-04-17 0.0
2012-04-18 0.0
2012-04-19 0.0
2012-04-20 0.0
Name: Forecast, dtype: float64

Continue reading...