Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology

ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

Forecasting Index of Industrial Production

Sub-Series Using Statistical and Deep
Learning Approach
P. Preethi1*; S. A. Jyothi Rani2; V. V. Haragopal3
1
Research Scholar, 2 Professor, 3Professor (Retd.)
1,2,3 Department of Statistics, Osmania University, Hyderabad, India

Corresponding Author: P. Preethi1*

Publication Date: 2025/08/13

Abstract: The Index of Industrial Production (IIP) is a key economic indicator that tracks manufacturing activity across
various sectors. This paper aims to predict the IIP for three sub-series—Mining, Manufacturing, and Electricity—using
both conventional statistical methods and deep learning approaches, analyzing data from April 2012 to September 2022.
Model performance is evaluated by comparing Mean Square Error (MSE), Mean Absolute Percentage Error (MAPE), and
Root Mean Square Error (RMSE)1. The results show that the RNN model outperforms other models for all three sub-series,
and is used to forecast these sub-series from October 2022 to September 2023.

Keywords: IIP, Mining, Manufacturing, Electricity, SARIMA, FFNN, RNN, LSTM, Forecasting.

How to Cite: P. Preethi; S. A. Jyothi Rani; V. V. Haragopal (2025) Forecasting Index of Industrial Production Sub-Series Using
Statistical and Deep Learning Approach. International Journal of Innovative Science and Research Technology,
10(7), 3539-3547. https://doi.org/10.38124/ijisrt/25jul1699

I. INTRODUCTION (ARIMA), Multilayer Perceptron (MLP), Recurrent Neural

Networks (RNN), and Convolutional Neural Networks
The Index of Industrial Production (IIP) is a key (CNN). The models are evaluated based on Mean Square
economic indicator that measures trends in industrial output Error (MSE), Root Mean Square Error (RMSE), and Mean
over time, based on a chosen base year. It reflects the relative Absolute Percentage Error (MAPE). The results indicate that
change in physical production in industries during a specified CNN outperforms ARIMA, RNN, and MLP in predicting rice
year compared to the previous year. The IIP covers three production3.
major sub-sectors: Mining, Manufacturing, and Electricity.
Policymakers, economists, analysts, and businesses closely Salam Shantikumar Singh, T. Loidang Devi, Tanusree
monitor these figures to assess industrial growth and inform Deb Roy (2016): “The industrial sector plays a crucial role in
decisions. This study aims to provide a reliable economic India's economic growth, with the composition and
indicator of industrial growth, using historical IIP data as a production of goods influenced by various factors. These
reference for future IIP releases. Statistical models like factors can cause short-term (seasonal) and long-term (trend)
ARIMA and deep learning models have been applied to the fluctuations in the Index of Industrial Production (IIP). The
IIP sub-series, and the best model is evaluated based on error primary objective of this study is to analyze the impact of
measures such as Mean Square Error (MSE), Mean Absolute seasonal trends on the IIP in India. Using data from the
Percentage Error (MAPE), and Root Mean Square Error National Data Sharing & Accessibility Policy (NDSAP),
(RMSE)2. Predictions are made for the next 12 months for the which includes the general IIP and 26 industrial sub-sectors,
Mining, Manufacturing, and Electricity sub-sectors using the the study employs the ARIMA (p, d, q) time series model to
best-performing model. assess these variations. The results indicate that both seasonal
and trend effects are present in the IIP, and a future forecast
II. LITERATURE REVIEW is made after adjusting for these fluctuations.” 4.

Rani, SA Jyothi, and N. Chandan Babu (2020): “This III. OBJECTIVES OF THE STUDY
paper presents the forecasting of rice production (in million
tonnes) using Auto Regressive Integrated Moving Averages  To fit the model using the ARIMA method

IJISRT25JUL1699 www.ijisrt.com 3539

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

 To fit the model using the Deep Learning Method-FFNN, “The results of forecasting are presented using
RNN, and LSTM, using python code. different methods. The methods are compared using Mean
Absolute Error (MAE), Root Mean Squared Error (RMSE),
 To Identify the best model by comparing MSE, MAE, and and Mean Absolute Percentage Error (MAPE) which are
RMSE out of all fitted SARIMA and Deep learning given below” [1]:
models.
MSE=1/n ∑ (Yt-Ft)2
 To forecast the IIP values for all three sub-series Mining,
Manufacturing, and Electricity for the next 12 months MAE=1/n ∑ |Yt-Ft|
using best-fit model.

IV. RESEARCH METHODOLOGY AND

RESULTS
Where Yt is the actual value, Ft is the fitted value and
 Data Source: n is the number of months used as forecasting period.
The study considers the Index of Industrial Production
data, published by the Ministry of Statistics and Programme  ARIMA Model:
Implementation (MOSPI). This study examines the above- “The development of the ARIMA model for a single
mentioned objectives at all Indian levels. The period of study variable involves identification, estimation, verification, and
is from April 2012 to September 2022 (A total of 126 forecasting. Each of these steps is now explained for all the
observations). IIP values over months for all three sub-series IIP’s three sub-series datasets” [1].
Mining, Manufacturing, and Electricity have been used.

Fig 1 Trends of IIP Data from April 2012 to Sept 2022 for all 3 Sectors

IJISRT25JUL1699 www.ijisrt.com 3540

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

“First examined whether the data is stationary or not the results show that the data is not stationary for all the sub-
using the Dickey-fuller Test for IIP’s three data series- series. Therefore, the data is converted into stationary by
Mining, Manufacturing, Electricity as shown in Table 1 and using the first or second-order difference of the datasets.

Table 1 Result of Dickey-Fuller Test

Dataset p-value Result
Mining 0.886 if above 0.05, data is not stationary
Manufacturing 0.623 if above 0.05, data is not stationary
Electricity 0.892 if above 0.05, data is not stationary

To prevent overfitting in neural networks, we divided In addition, models are fitted using different
the original dataset into training and testing sets. We tested specifications as described below.
different combinations of these datasets across the following
neural network architectures:  Activation function: Rectified Linear Unit;
 Validation Generator: Timeseries Generator;
Feed Forward with splits of [70%, 30%], [75%, 25%],  Batch and Sequence Size: 1 and 12.
and [80%, 20%]; RNN with splits of [70%, 30%], [75%,
25%], and [80%, 20%]; and LSTM with splits of [70%, 30%], The following three metrics— MSE, MAE, and
[75%, 25%], and [80%, 20%]. These analyses were RMSE, are used to compare the methods.
conducted across three different data series: Mining,
Manufacturing, and Electricity for the training datasets and
the results are presented in Table 2.

Table 2 Test Accuracy: (Training Dataset)

Mining Manufacturing Electricity
Model Name Specification Training MSE MAE RMSE MSE MAE RMSE MSE MAE RMSE
Dataset
Feed DL (64) 70% 5.61 1.84 2.37 21.31 3.71 4.61 18.95 3.52 4.35
Forward DL (32)
DL (1)
Feed DL (64) 75% 10.54 2.62 3.24 21.43 3.62 4.62 26.05 4.13 5.10
Forward DL (32)
DL (1)
Feed DL (64) 80% 16.92 2.83 4.11 95.25 4.87 9.75 33.63 4.45 5.79
Forward DL (32)
DL (1)
RNN Simple RNN 70% 6.73 2.10 2.59 9.97 2.58 3.15 9.72 2.37 3.11
(64)
DL (1)
RNN Simple RNN 75% 3.02 1.40 1.73 7.48 2.19 2.73 6.12 1.90 2.47
(64)
DL (1)
RNN Simple RNN 80% 6.21 2.04 2.49 60.43 4.39 7.77 6.46 1.81 2.54
(64)
DL (1)

RNN Simple RNN 70% 7.95 2.30 2.82 7.58 2.25 2.75 13.90 3.07 3.72
(64)
DL (32)
DL (1)
RNN Simple RNN 75% 3.72 1.53 1.92 15.62 3.20 3.95 6.14 1.88 2.47
(64)
DL (32)
DL (1)
RNN Simple RNN 80% 3.61 1.50 1.90 58.13 4.99 7.62 7.83 2.12 2.79
(64)
DL (32)
DL (1)
LSTM LSTM (50) 70% 24.60 3.86 4.95 13.29 2.86 3.64 16.88 3.36 4.10

IJISRT25JUL1699 www.ijisrt.com 3541

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

LSTM (50)
DL (1)
LSTM LSTM (50) 75% 15.54 3.18 3.94 34.12 4.91 5.84 23.14 4.01 4.81
LSTM (50)
DL (1)
LSTM LSTM (50) 80% 6.21 2.04 2.49 89.04 4.80 9.43 32.23 4.03 5.67
LSTM (50)
DL (1)
LSTM LSTM (50) 70% 16.05 3.13 4.00 14.37 3.15 3.79 33.80 4.77 5.81
LSTM (50)
DL (32)
DL (1)
LSTM LSTM (50) 75% 20.47 3.54 4.52 23.99 3.84 4.89 44.76 5.46 6.69
LSTM (50)
DL (32)
DL (1)
LSTM LSTM (50) 80% 17.22 3.06 4.15 81.02 4.87 9.00 32.04 4.27 5.66
LSTM (50)
DL (32)
DL (1)

Referring to Table 2, it becomes evident that the MSE, models across all series, namely Mining, Manufacturing, and
MAE, and RMSE values are consistently minimized when Electricity. With the adoption of the 75% training dataset,
75% of the training observations are used. Consequently, it SARIMA models were implemented for all three data series
was decided that 75% of the observations would serve as the -Mining, Manufacturing, and Electricity which are shown in
training dataset, while the remaining 25% would be Tables 3, 4, and 5.
designated as the testing dataset for the execution of different

Table 3 Mining-ARIMA Best Model Identification Results

IJISRT25JUL1699 www.ijisrt.com 3542

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

Table 4 Manufacturing-ARIMA Best Model Identification Results

Table 5 Electricity-ARIMA Best Model Identification Results

IJISRT25JUL1699 www.ijisrt.com 3543

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

The best models are Mining-SARIMA (1,1,1) (0,1,1) FFNN, RNN, and LSTM are executed using the commands
[12], Manufacturing- SARIMA (0,1,1) (1,1,1) [12], called Sequential, Simple RNN, and LSTM in Python. The
Electricity-SARIMA (1,1,0) (0,1,1) [12]. Actual Vs. Fitted values have been plotted for the Mining
series for all the models with [75%, 25%] split shown in Fig.
After fitting the SARIMA model, the other models 2. The blue trend indicates the mining IIP values, the orange
FFNN, RNN, and LSTM were fitted by considering training trend indicates the predicted training values and the green
and testing split of [75%,25%] using Python code. The trend indicates the predicted testing values.

Fig 2 Actual vs Fitted Values for Mining Series

(Both Training and Testing Have Been Plotted for all the Models i.e. ARIMA, FFNN, RNN, and LSTM)

The above Fig. 2, clearly shows that the SARIMA, the accuracy of projection is also quantified for the testing
FFNN, and RNN model fits well for training datasets whereas dataset by calculating the MSE, MAE, and RMSE of the
RNN and LSTM models fit well for testing data. Therefore, testing dataset of the mining series.

Table 6 Mining: Test Accuracy

Model Name Specification Testing Data Set MSE MAE RMSE
SARIMA - 25% 137.92 9.14 11.74
Feed Forward Dense Layer-64 25% 38.72 4.82 6.22
Dense Layer-32
Dense Layer-1
RNN Simple RNN-64 Dense Layer-1 25% 24.47 4.20 4.94
LSTM LSTM (50) LSTM (50) Dense Layer-1 25% 78.68 7.83 8.87

IJISRT25JUL1699 www.ijisrt.com 3544

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

From above Table 6, we observe that MSE, MAE, and Similarly, The Actual Vs. Fitted values have been
RMSE are the least for the RNN model. Therefore, from plotted for the Manufacturing series for all the models with
Table 6 and Table 2, we conclude that RNN (SimpleRNN-64, [75%, 25%] split shown in Fig. 3. The blue trend indicates the
Dense Layer-1) is the best fit for testing as well as training manufacturing IIP values, the orange trend indicates the
datasets of mining data series. Hence, we conclude that RNN predicted training values and the green trend indicates the
(SimpleRNN-64, Dense Layer-1) model outperforms the predicted testing values.
other models and will be used for forecasting the mining data
series.

Fig 3 Actual vs Fitted Values for Manufacturing Series

(Both Training and Testing Have Been Plotted for all the Models i.e. ARIMA, FFNN, RNN, and LSTM)

The above Fig 3 clearly shows that the SARIMA, RNN projection is also quantified for the testing dataset by
model fits well for training datasets whereas only the RNN calculating MSE, MAE, and RMSE to get the accuracy of the
model fits well for testing data. Therefore, the accuracy of testing dataset of the manufacturing series.

Table 7 Manufacturing-Test Accuracy

Model Name Specification Testing Data Set MSE MAE RMSE
SARIMA - 25% 137.92 9.14 11.74
Feed Forward Dense Layer-64 25% 38.72 4.82 6.22
Dense Layer-32
Dense Layer-1
RNN Simple RNN-64 Dense Layer-1 25% 24.47 4.20 4.94
LSTM LSTM (50) LSTM (50) Dense Layer-1 25% 78.68 7.83 8.87

IJISRT25JUL1699 www.ijisrt.com 3545

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

From Table 7 we observe that MSE, MAE, and RMSE the manufacturing data series. Similarly, The Actual Vs.
are the least for the RNN model. Therefore, from Table 7 and Fitted values have been plotted for the Electricity series for
Table 2, we conclude that RNN (SimpleRNN-64, Dense all the models with [75%, 25%] split shown in Fig 4. The blue
Layer-1) is the best fit for testing as well as the training trend indicates the electricity IIP values, the orange trend
dataset of the Manufacturing data series. Hence, we conclude indicates the predicted training values and the green trend
that the RNN (SimpleRNN-64, Dense Layer-1) model indicates the predicted testing values.
outperforms the other models and will be used for forecasting

Fig 4 Actual vs Fitted Values for Electricity Series

(Both Training and Testing Have Been Plotted for all the Models i.e. ARIMA, FFNN, RNN, and LSTM)

The above Fig. 4 clearly shows that the SARIMA, accuracy of projection is also quantified for the testing dataset
FNN, and RNN model fits well for training datasets whereas by calculating MSE, MAE, and RMSE of the testing dataset
only the RNN model fits well for testing data. Therefore, the of the electricity series.

IJISRT25JUL1699 www.ijisrt.com 3546

Volume 10, Issue 7, July – 2025 International Journal of Innovative Science and Research Technology
ISSN No: 2456-2165 https://doi.org/10.38124/ijisrt/25jul1699

Table 8 Electricity-Test Accuracy

Model Name Specification Testing Data Set MSE MAE RMSE
SARIMA - 25% 137.92 9.14 11.74
Feed Forward Dense Layer-64 25% 38.72 4.82 6.22
Dense Layer-32
Dense Layer-1
RNN Simple RNN-64 Dense Layer-1 25% 24.47 4.20 4.94
LSTM LSTM (50) LSTM (50) Dense Layer-1 25% 78.68 7.83 8.87

From the above Table 8 we observe that MSE, MAE, V. CONCLUSION

and RMSE are the least for the RNN model. Therefore, from
Table 8 and Table 2, we conclude that RNN (SimpleRNN-64, According to the findings, RNN outperforms all other
Dense Layer-1) is the best fit for testing as well as training models for all the data series - Mining, Manufacturing, and
datasets of the Electricity data series. Hence, we conclude that Electricity. Therefore, forecasted values for all the series-
the RNN (SimpleRNN-64, Dense Layer-1) model Mining Manufacturing and Electricity are projected using the
outperforms the other models and will be used for forecasting RNN (SimpleRNN-64, Dense Layer-1) model which is
the electricity data series. shown below in Table 9.

Table 9 Forecasted Values of IIP from Oct-2020 to Sept-2023 for all three Series
Months Mining Manufacturing Electricity
Oct-22 102.66 127.61 191.08
Nov-22 103.13 127.73 197.19
Dec-22 117.22 127.35 196.83
Jan-23 128.72 129.72 194.27
Feb-23 133.43 124.64 189.90
Mar-23 146.68 132.74 183.75
Apr-23 162.12 124.34 181.36
May-23 171.21 131.89 183.81
Jun-23 183.95 126.58 191.10
Jul-23 177.63 130.28 199.06
Aug-23 172.84 126.90 204.47
Sep-23 160.75 129.01 202.47

REFERENCES

[1]. Patil, Preethi, S. A. Jyothirani, and V. V. Haragopal.

"Impact of Lockdown on India's Index of Industrial
Production–Traditional and Deep Learning Statistical
Approach." European Journal of Mathematics and
Statistics 3.4 (2022): 62-70.
[2]. Sodhi, ManMohan S., et al. "A robust and forward-
Looking industrial production indicator." Economic
and Political Weekly (2013): 126-130.
[3]. Rani, SA Jyothi, and N. Chandan Babu. "Forecasting
production of rice in India–using Arima and deep
learning methods." Int J Math Trends Technol (IJMTT)
66.4 (2020).
[4]. Singh, Salam Shantikumar, T. Loidang Devi, and T.
Deb Roy. "Time series analysis of the index of
industrial production of India." IOSR Journal of
Mathematics 12.3 (2016): 1-7.

IJISRT25JUL1699 www.ijisrt.com 3547