1/3/24, 6:10 PM PPP Models - GARCH & NARNN.

ipynb - Colaboratory

keyboard_arrow_down Petroleum Price Prediction Models

The following models would be developed:

GARCH with NARNN and vice versa

ARIMA with NARNN vice versa

SARIMA with NARNN vice versa

Exponential Smoothing with NARNN and vice versa.

NB: The residuals from the base model would be used to train the other model

The existing models serve as a Benchmark ( GARCH, ARIMA and SARIMA, Exponential
Smoothing)

1 # Load Libraries
2 import numpy as np # Mathematical Computations
3 import pandas as pd # Data Manipulation
4 import matplotlib.pyplot as plt # Data Visualization
5 import seaborn as sns # Advance Data Visualizations
6 import plotly.express as px # Advance and Interactive Data Visualizations

keyboard_arrow_down Load Dataset and Preprocessing

1 # Pick the data from local machine

2 from google.colab import files
3 file_path = files.upload()

Choose Files petroleum_products.xlsx

petroleum_products.xlsx(application/vnd.openxmlformats-officedocument.spreadsheetml.sheet) -
10200 bytes, last modified: 12/21/2023 - 100% done
Saving petroleum_products.xlsx to petroleum_products.xlsx

1 # Load the dataset into a useable dataframe format

2 df = pd.read_excel('petroleum_products.xlsx', index_col='Date')
3 df.head()

GASOLINE DIESEL LPG

Date

2019-01-01 4.89 4.90 4.99

2019-02-01 4.90 4.91 4.94

2019-03-01 5.15 5.16 5.01

2019-04-01 5.19 5.19 5.39

2019-05-01 5.20 5.21 4.97

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 1/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
1 # Rename columns for convinience
2 df = df.rename(columns = {'GASOLINE': 'Gasoline', 'DIESEL': 'Diesel'})
3 df.columns

Index(['Gasoline', 'Diesel', 'LPG'], dtype='object')

keyboard_arrow_down Data Selection

1 # Select a Product and Model Type to Analyze

product: Gasoline
2 product = "Gasoline" # @param ["Gasoline", "Die
3 print(f"Selected Product: {product}")

Selected Product: Gasoline

keyboard_arrow_down GARCH - NARNN Model

keyboard_arrow_down Estimate The Mean Equation

1 ! pip install pmdarima

2 import pmdarima as pm
3 import statsmodels.api as sm
4 import matplotlib.pyplot as plt
5 from statsmodels.graphics.tsaplots import plot_acf, plot_pacf

Collecting pmdarima
Downloading pmdarima-2.0.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.manyl
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2.1/2.1 MB 9.5 MB/s eta 0:00:00
Requirement already satisfied: joblib>=0.11 in /usr/local/lib/python3.10/dist-packages (f
Requirement already satisfied: Cython!=0.29.18,!=0.29.31,>=0.29 in /usr/local/lib/python3
Requirement already satisfied: numpy>=1.21.2 in /usr/local/lib/python3.10/dist-packages (
Requirement already satisfied: pandas>=0.19 in /usr/local/lib/python3.10/dist-packages (f
Requirement already satisfied: scikit-learn>=0.22 in /usr/local/lib/python3.10/dist-packa
Requirement already satisfied: scipy>=1.3.2 in /usr/local/lib/python3.10/dist-packages (f
Requirement already satisfied: statsmodels>=0.13.2 in /usr/local/lib/python3.10/dist-pack
Requirement already satisfied: urllib3 in /usr/local/lib/python3.10/dist-packages (from p
Requirement already satisfied: setuptools!=50.0.0,>=38.6.0 in /usr/local/lib/python3.10/d
Requirement already satisfied: packaging>=17.1 in /usr/local/lib/python3.10/dist-packages
Requirement already satisfied: python-dateutil>=2.8.1 in /usr/local/lib/python3.10/dist-p
Requirement already satisfied: pytz>=2020.1 in /usr/local/lib/python3.10/dist-packages (f
Requirement already satisfied: threadpoolctl>=2.0.0 in /usr/local/lib/python3.10/dist-pac
Requirement already satisfied: patsy>=0.5.4 in /usr/local/lib/python3.10/dist-packages (f
Requirement already satisfied: six in /usr/local/lib/python3.10/dist-packages (from patsy
Installing collected packages: pmdarima
Successfully installed pmdarima-2.0.4

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 2/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
1 from statsmodels.tsa.stattools import adfuller
2 # Perform the ADF test
3 def perform_adf_test(ds):
4 result = adfuller(ds)
5 # Extract and print test statistics
6 adf_statistic = result[0]
7 p_value = result[1]
8 critical_values = result[4]
9
10 print(f'ADF Statistic: {adf_statistic}')
11 print(f'p-value: {p_value}')
12 print(f'Critical Values: {critical_values}')
13
14 # Check the p-value against the significance level (e.g., 0.05)
15 if p_value <= 0.05:
16 print("Reject the null hypothesis. The time series is likely stationary.")
17 else:
18 print("Fail to reject the null hypothesis. The time series may not be stationary.")
19
20 perform_adf_test(df[product])

ADF Statistic: -0.4714985407350731

p-value: 0.8974593895763701
Critical Values: {'1%': -3.5506699942762414, '5%': -2.913766394626147, '10%': -2.59462404
Fail to reject the null hypothesis. The time series may not be stationary.

1 # Diffrence the data

2 differenced_ds = df[product].diff(1).dropna()
3 perform_adf_test(differenced_ds)

ADF Statistic: -6.54915636272955

p-value: 8.941192716390255e-09
Critical Values: {'1%': -3.5506699942762414, '5%': -2.913766394626147, '10%': -2.59462404
Reject the null hypothesis. The time series is likely stationary.

1 def find_arima_lags(data, max_ar=5, max_ma=5, figsize=(12, 3)):

2 # Plot ACF and PACF
3 fig, (ax1, ax2) = plt.subplots(1, 2, figsize=figsize)
4
5 plot_acf(data, ax=ax1)
6 plot_pacf(data, ax=ax2)
7
8 ax1.set_title('Autocorrelation Function (ACF)')
9 ax2.set_title('Partial Autocorrelation Function (PACF)')
10
11 plt.show()
12
13 find_arima_lags(differenced_ds)

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 3/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory

1 # Perform Stepwise fit to identify the best order

2 stepwise_fit = pm.auto_arima(df[product], trace=True)
3 print(stepwise_fit.summary())

Performing stepwise search to minimize aic

ARIMA(2,1,2)(0,0,0)[0] intercept : AIC=154.005, Time=0.34 sec
ARIMA(0,1,0)(0,0,0)[0] intercept : AIC=153.281, Time=0.06 sec
ARIMA(1,1,0)(0,0,0)[0] intercept : AIC=154.166, Time=0.05 sec
ARIMA(0,1,1)(0,0,0)[0] intercept : AIC=152.760, Time=0.08 sec
ARIMA(0,1,0)(0,0,0)[0] : AIC=152.667, Time=0.03 sec
ARIMA(1,1,1)(0,0,0)[0] intercept : AIC=151.680, Time=0.33 sec
ARIMA(2,1,1)(0,0,0)[0] intercept : AIC=152.396, Time=0.45 sec
ARIMA(1,1,2)(0,0,0)[0] intercept : AIC=152.019, Time=0.27 sec
ARIMA(0,1,2)(0,0,0)[0] intercept : AIC=150.020, Time=0.46 sec
ARIMA(0,1,3)(0,0,0)[0] intercept : AIC=152.019, Time=0.56 sec
ARIMA(1,1,3)(0,0,0)[0] intercept : AIC=154.020, Time=0.55 sec
ARIMA(0,1,2)(0,0,0)[0] : AIC=149.611, Time=0.22 sec
ARIMA(0,1,1)(0,0,0)[0] : AIC=151.556, Time=0.11 sec
ARIMA(1,1,2)(0,0,0)[0] : AIC=151.592, Time=0.38 sec
ARIMA(0,1,3)(0,0,0)[0] : AIC=151.588, Time=0.25 sec
ARIMA(1,1,1)(0,0,0)[0] : AIC=150.689, Time=0.31 sec
ARIMA(1,1,3)(0,0,0)[0] : AIC=152.953, Time=0.80 sec

Best model: ARIMA(0,1,2)(0,0,0)[0]

Total fit time: 5.355 seconds
SARIMAX Results
==============================================================================
Dep. Variable: y No. Observations: 60
Model: SARIMAX(0, 1, 2) Log Likelihood -71.805
Date: Wed, 03 Jan 2024 AIC 149.611
Time: 17:31:45 BIC 155.844
Sample: 01-01-2019 HQIC 152.044
- 12-01-2023
Covariance Type: opg
==============================================================================
coef std err z P>|z| [0.025 0.975]
------------------------------------------------------------------------------
ma.L1 0.2512 0.068 3.721 0.000 0.119 0.384
ma.L2 -0.2579 0.151 -1.705 0.088 -0.554 0.039
sigma2 0.6648 0.042 15.793 0.000 0.582 0.747
===================================================================================
Ljung-Box (L1) (Q): 0.05 Jarque-Bera (JB): 1182.36
Prob(Q): 0.82 Prob(JB): 0.00
Heteroskedasticity (H): 23.10 Skew: 3.51
Prob(H) (two-sided): 0.00 Kurtosis: 23.77
===================================================================================

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 4/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
Warnings:
[1] Covariance matrix calculated using the outer product of gradients (complex-step).

1 from statsmodels.tsa.arima.model import ARIMA

2 # # Fit ARIMA(2,2) model
3 order = (0, 1, 2) # (p, d, q) order
4 arima_model = ARIMA(df[product], order=order)
5 arima_results = arima_model.fit()
6
7 # Display model summary
8 print(arima_results.summary())

SARIMAX Results
==============================================================================
Dep. Variable: Gasoline No. Observations: 60
Model: ARIMA(0, 1, 2) Log Likelihood -71.805
Date: Wed, 03 Jan 2024 AIC 149.611
Time: 17:32:54 BIC 155.844
Sample: 01-01-2019 HQIC 152.044
- 12-01-2023
Covariance Type: opg
==============================================================================
coef std err z P>|z| [0.025 0.975]
------------------------------------------------------------------------------
ma.L1 0.2512 0.068 3.721 0.000 0.119 0.384
ma.L2 -0.2579 0.151 -1.705 0.088 -0.554 0.039
sigma2 0.6648 0.042 15.793 0.000 0.582 0.747
===================================================================================
Ljung-Box (L1) (Q): 0.05 Jarque-Bera (JB): 1182.36
Prob(Q): 0.82 Prob(JB): 0.00
Heteroskedasticity (H): 23.10 Skew: 3.51
Prob(H) (two-sided): 0.00 Kurtosis: 23.77
===================================================================================

Warnings:
[1] Covariance matrix calculated using the outer product of gradients (complex-step).
/usr/local/lib/python3.10/dist-packages/statsmodels/tsa/base/tsa_model.py:473: ValueWarni
self._init_dates(dates, freq)
/usr/local/lib/python3.10/dist-packages/statsmodels/tsa/base/tsa_model.py:473: ValueWarni
self._init_dates(dates, freq)
/usr/local/lib/python3.10/dist-packages/statsmodels/tsa/base/tsa_model.py:473: ValueWarni
self._init_dates(dates, freq)

1 arima_resid = pd.DataFrame(arima_results.resid)
2 arima_resid.plot(kind = "kde", figsize=(20,6))

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 5/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory

<Axes: ylabel='Density'>

keyboard_arrow_down ARCH TEST

Since the data has an AR component, we would fit the ARCH Model on the residuals of the ARIMA
model.

1 ! pip install arch

2 import arch

Requirement already satisfied: arch in /usr/local/lib/python3.10/dist-packages (6.2.0)

Requirement already satisfied: numpy>=1.19 in /usr/local/lib/python3.10/dist-packages (fr
Requirement already satisfied: scipy>=1.5 in /usr/local/lib/python3.10/dist-packages (fro
Requirement already satisfied: pandas>=1.1 in /usr/local/lib/python3.10/dist-packages (fr
Requirement already satisfied: statsmodels>=0.12 in /usr/local/lib/python3.10/dist-packag
Requirement already satisfied: python-dateutil>=2.8.1 in /usr/local/lib/python3.10/dist-p
Requirement already satisfied: pytz>=2020.1 in /usr/local/lib/python3.10/dist-packages (f
Requirement already satisfied: patsy>=0.5.4 in /usr/local/lib/python3.10/dist-packages (f
Requirement already satisfied: packaging>=21.3 in /usr/local/lib/python3.10/dist-packages
Requirement already satisfied: six in /usr/local/lib/python3.10/dist-packages (from patsy

1 # Obtain squared residuals

2 # squared_residuals = arima_results.resid**2
3 ds = arima_resid #df[product]
4 lags = 30 # ARCH TEST IS POSITIVE AT LAG 30
5
6 # Fit an ARCH model to the squared residuals
7 arch_model = arch.arch_model(ds, mean='Zero', vol='ARCH', p=3, lags=lags)
8 arch_result = arch_model.fit()
9
10 print(arch_result.summary())
11
12 # Perform ARCH LM test
13 lm_test = arch_result.arch_lm_test(lags=lags) # Adjust the number of lags as needed
14 print("ARCH LM Test p-value:", lm_test)
15

Iteration: 1, Func. Count: 6, Neg. LLF: 130.25413733733095

Iteration: 2, Func. Count: 16, Neg. LLF: 2024.039104072993
Iteration: 3, Func. Count: 22, Neg. LLF: 1953.3446359373388
Iteration: 4, Func. Count: 28, Neg. LLF: 1830.1700243650791
Iteration: 5, Func. Count: 34, Neg. LLF: 1772.5178015616962
Iteration: 6, Func. Count: 40, Neg. LLF: 1772.5177631819702

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 6/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
Iteration: 7, Func. Count: 46, Neg. LLF: 1772.517900132466
Iteration: 8, Func. Count: 52, Neg. LLF: 57.32423354514197
Iteration: 9, Func. Count: 57, Neg. LLF: 57.46116408610714
Iteration: 10, Func. Count: 63, Neg. LLF: 57.28839336266442
Iteration: 11, Func. Count: 68, Neg. LLF: 1772.5140329794292
Iteration: 12, Func. Count: 75, Neg. LLF: 57.288380645096915
Optimization terminated successfully (Exit mode 0)
Current function value: 57.28837974861303
Iterations: 13
Function evaluations: 75
Gradient evaluations: 12
Zero Mean - ARCH Model Results
==============================================================================
Dep. Variable: 0 R-squared: 0.000
Mean Model: Zero Mean Adj. R-squared: 0.017
Vol Model: ARCH Log-Likelihood: -57.2884
Distribution: Normal AIC: 122.577
Method: Maximum Likelihood BIC: 130.954
No. Observations: 60
Date: Wed, Jan 03 2024 Df Residuals: 60
Time: 17:59:52 Df Model: 0
Volatility Model
=============================================================================
coef std err t P>|t| 95.0% Conf. Int.
-----------------------------------------------------------------------------
omega 0.1130 4.473e-02 2.527 1.149e-02 [2.538e-02, 0.201]
alpha[1] 1.0000 0.396 2.522 1.166e-02 [ 0.223, 1.777]
alpha[2] 6.5873e-14 2.325e-02 2.833e-12 1.000 [-4.557e-02,4.557e-02]
alpha[3] 1.0795e-13 8.136e-03 1.327e-11 1.000 [-1.595e-02,1.595e-02]
=============================================================================

Covariance estimator: robust

ARCH LM Test p-value: ARCH-LM Test
H0: Residuals are homoskedastic.
ARCH-LM Test
H1: Residuals are conditionally heteroskedastic.
Statistic: 49.3932
P-value: 0.0105
Distributed: chi2(29)

1 # Create a GARCH model

2 # Extract Residuals From ARCH Model
3 arch_residuals = arch_result.resid
4 garch_model = arch.arch_model(arch_residuals, mean='Constant', lags=1, vol='GARCH', p=1, q=1, power=2.0,
5
6 # Fit the model
7 garch_result = garch_model.fit()
8
9 # Display model summary
10 print(garch_result.summary())

Iteration: 1, Func. Count: 6, Neg. LLF: 248.13952546070539

Iteration: 2, Func. Count: 13, Neg. LLF: 140.93626886556575
Iteration: 3, Func. Count: 19, Neg. LLF: 122.20202445832632
Iteration: 4, Func. Count: 25, Neg. LLF: 123.06109763174514
Iteration: 5, Func. Count: 31, Neg. LLF: 3552.87176654639
Iteration: 6, Func. Count: 37, Neg. LLF: 8905.56417193098
Iteration: 7, Func. Count: 43, Neg. LLF: 602.6142582806667
Iteration: 8, Func. Count: 49, Neg. LLF: 6475.377231855412
Iteration: 9, Func. Count: 55, Neg. LLF: 5080.56075662095
Iteration: 10, Func. Count: 61, Neg. LLF: 58.103062449764415
Iteration: 11, Func. Count: 66, Neg. LLF: 385262.8141300336

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 7/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
Iteration: 12, Func. Count: 72, Neg. LLF: 56.860447795643275
Iteration: 13, Func. Count: 77, Neg. LLF: 56.81596564847961
Iteration: 14, Func. Count: 82, Neg. LLF: 56.8117858529232
Iteration: 15, Func. Count: 87, Neg. LLF: 56.8116901249659
Iteration: 16, Func. Count: 92, Neg. LLF: 59.88033886994014
Iteration: 17, Func. Count: 100, Neg. LLF: 56.91129629217719
Optimization terminated successfully (Exit mode 0)
Current function value: 56.81168551834978
Iterations: 18
Function evaluations: 103
Gradient evaluations: 17
Constant Mean - GARCH Model Results
==============================================================================
Dep. Variable: resid R-squared: 0.000
Mean Model: Constant Mean Adj. R-squared: 0.000
Vol Model: GARCH Log-Likelihood: -56.8117
Distribution: Normal AIC: 121.623
Method: Maximum Likelihood BIC: 130.001
No. Observations: 60
Date: Wed, Jan 03 2024 Df Residuals: 59
Time: 18:04:45 Df Model: 1
Mean Model
===========================================================================
coef std err t P>|t| 95.0% Conf. Int.
---------------------------------------------------------------------------
mu 0.0678 4.784e-02 1.417 0.156 [-2.597e-02, 0.162]
Volatility Model
=============================================================================
coef std err t P>|t| 95.0% Conf. Int.
-----------------------------------------------------------------------------
omega 0.1003 3.882e-02 2.584 9.771e-03 [2.422e-02, 0.176]
alpha[1] 1.0000 0.366 2.730 6.327e-03 [ 0.282, 1.718]
beta[1] 1.5508e-13 8.794e-03 1.763e-11 1.000 [-1.724e-02,1.724e-02]
=============================================================================

Covariance estimator: robust

keyboard_arrow_down Evaluate GARCH Model

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 8/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
1 # prompt: Write codes to evaluate the GARCH Model
2
3 # Evaluate the GARCH Model
4 # Calculate Residuals
5 residuals = garch_result.resid
6
7 # Calculate the Root Mean Squared Error (RMSE)
8 rmse = np.sqrt(np.mean(residuals**2))
9 print(f"RMSE: {rmse}")
10
11 # Calculate the Mean Absolute Error (MAE)
12 mae = np.mean(np.abs(residuals))
13 print(f"MAE: {mae}")
14
15 # Calculate the Mean Absolute Percentage Error (MAPE)
16 mape = np.mean(np.abs(residuals / arch_residuals)) * 100
17 print(f"MAPE: {mape}%")
18
19 # Plot the Residuals
20 plt.figure(figsize=(12, 6))
21 plt.plot(residuals, label='Residuals', alpha=0.6)
22 plt.title(f'Residuals of GARCH Model for {product}')
23 plt.xlabel('Year')
24 plt.ylabel('Residuals')
25 plt.legend()
26 plt.show()
27

RMSE: 1.0142621472049587
MAE: 0.4755606334037595
MAPE: 145.18138560695348%

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 9/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
1 import statsmodels.api as sm
2 from statsmodels.stats.diagnostic import acorr_ljungbox
3
4 # Residual Analysis
5 squared_residuals = residuals**2
6
7 # Calculate standardized squared residuals
8 standardized_squared_residuals = (residuals ** 2) / (garch_result.conditional_volatility ** 2)
9
10
11 # Diagnostic Tests
12 # Autocorrelation in squared residuals (Ljung-Box test)
13 print(acorr_ljungbox(standardized_squared_residuals, lags=10, boxpierce=True))

lb_stat lb_pvalue bp_stat bp_pvalue

1 0.018688 0.891263 0.017784 0.893911
2 0.137440 0.933588 0.128875 0.937595
3 0.391232 0.942048 0.362199 0.947936
4 0.518298 0.971695 0.476969 0.975704
5 0.547749 0.990270 0.503095 0.992010
6 0.575410 0.996796 0.527186 0.997492
7 0.916299 0.996068 0.818591 0.997250
8 0.998753 0.998256 0.887746 0.998863
9 1.453216 0.997479 1.261578 0.998559
10 1.453267 0.999074 1.261620 0.999506

1 # Generate GARCH predictions

2 garch_forecast = garch_result.forecast(reindex=False, start=12) # Replace with the desired start date
3
4 # Extract actual variances from the dataset
5 actual_variances = garch_result.conditional_volatility
6
7
8 # Extract the predicted conditional volatility
9 predicted_volatility = garch_forecast.variance['h.1']
10
11 # Plot the original returns and predicted volatility
12 plt.figure(figsize=(12, 6))
13 # plt.plot(actual_variances, label=f'Original {product}', alpha=0.6)
14 plt.plot(predicted_volatility.index, predicted_volatility, label='Predicted Volatility', linestyle='--',
15 plt.title('GARCH Model Volatility Forecast (In sample)')
16 plt.xlabel('Time')
17 plt.ylabel('Volatility')
18 plt.legend()
19 plt.show()
20

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 10/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory

Extract GARCH Residuals

1 # Extract residuals
2 garch_residuals = garch_result.resid
3 # Plot residuals
4 plt.figure(figsize=(10, 4))
5 plt.plot(garch_residuals, label='GARCH Residuals')
6 plt.title('Residuals of GARCH Model')
7 plt.xlabel('Time')
8 plt.ylabel('Residuals')
9 plt.legend()
10 plt.show()

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 11/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory

NARNN MODEL
keyboard_arrow_down

1 from tensorflow import keras

2 from keras.models import Sequential
3 from keras.layers import LSTM, Dense
4 from keras.optimizers import Adam
5 from sklearn.preprocessing import MinMaxScaler

1 # Extract residuals from the GARCH model

2 residuals = garch_result.resid
3
4 # Scale the residuals between 0 and 1
5 # scaler = MinMaxScaler(feature_range=(0, 1))
6 residuals_scaled = residuals.values.reshape(-1, 1) #scaler.fit_transform()
7
8 # Create a NARNN model
9 model = Sequential()
10 model.add(LSTM(units=100, activation='relu', return_sequences=True, input_shape=(residuals_scaled.shape[1
11 model.add(LSTM(units=100, activation='relu'))
12 model.add(Dense(units=1))
13
14 # Compile the model
15 model.compile(optimizer=Adam(learning_rate=0.001), loss='mean_squared_error')
16
17 # Train the model
18 model.fit(residuals_scaled, residuals_scaled, epochs=100, batch_size=32)

Epoch 1/100
2/2 [==============================] - 10s 42ms/step - loss: 1.0298
Epoch 2/100
2/2 [==============================] - 0s 24ms/step - loss: 1.0240
Epoch 3/100
2/2 [==============================] - 0s 26ms/step - loss: 1.0166
Epoch 4/100
2/2 [==============================] - 0s 20ms/step - loss: 1.0096
Epoch 5/100
2/2 [==============================] - 0s 16ms/step - loss: 1.0060
Epoch 6/100
2/2 [==============================] - 0s 25ms/step - loss: 0.9984
Epoch 7/100
2/2 [==============================] - 0s 27ms/step - loss: 0.9935
Epoch 8/100
2/2 [==============================] - 0s 16ms/step - loss: 0.9872
Epoch 9/100
2/2 [==============================] - 0s 17ms/step - loss: 0.9807
Epoch 10/100
2/2 [==============================] - 0s 23ms/step - loss: 0.9744
Epoch 11/100
2/2 [==============================] - 0s 19ms/step - loss: 0.9664
Epoch 12/100
2/2 [==============================] - 0s 29ms/step - loss: 0.9567
Epoch 13/100
2/2 [==============================] - 0s 19ms/step - loss: 0.9490
Epoch 14/100
2/2 [==============================] - 0s 21ms/step - loss: 0.9355
Epoch 15/100
2/2 [==============================] - 0s 20ms/step - loss: 0.9262
Epoch 16/100
2/2 [==============================] - 0s 23ms/step - loss: 0.9164
Epoch 17/100

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 12/13
1/3/24, 6:10 PM PPP Models - GARCH & NARNN.ipynb - Colaboratory
2/2 [==============================] - 0s 20ms/step - loss: 0.8984
Epoch 18/100
2/2 [==============================] - 0s 11ms/step - loss: 0.8840
Epoch 19/100
2/2 [==============================] - 0s 10ms/step - loss: 0.8729
Epoch 20/100
2/2 [==============================] - 0s 10ms/step - loss: 0.8499
Epoch 21/100
2/2 [==============================] - 0s 12ms/step - loss: 0.8302
Epoch 22/100
2/2 [==============================] - 0s 12ms/step - loss: 0.8135
Epoch 23/100
2/2 [==============================] - 0s 10ms/step - loss: 0.7892
Epoch 24/100
2/2 [==============================] - 0s 11ms/step - loss: 0.7639
Epoch 25/100
2/2 [==============================] - 0s 10ms/step - loss: 0.7336
Epoch 26/100
2/2 [==============================] - 0s 10ms/step - loss: 0.7076
Epoch 27/100
2/2 [==============================] - 0s 18ms/step - loss: 0.6580
Epoch 28/100
2/2 [==============================] - 0s 11ms/step - loss: 0.6259
Epoch 29/100
2/2 [==============================] - 0s 10ms/step - loss: 0.5784

1 # Evaluate the model

2 model.evaluate(residuals_scaled, residuals_scaled)

2/2 [==============================] - 0s 8ms/step - loss: 0.0016

0.001564527628943324

1 # Make predictions
2 predictions = model.predict(residuals_scaled)
3
4 # Rescale the predictions back to the original range
5 #predictions = scaler.inverse_transform(predictions)
6
7 df2 = df

https://colab.research.google.com/drive/12zwsFRYplo5DU6lIq6vfaYs75v6Sez4M#scrollTo=xFe4ZkpeBp26&uniqifier=3&printMode=true 13/13