| Title: | Neural Network–Based Model Selection and Forecasting |
| Type: | Package |
| Version: | 0.1.0 |
| Author: | Dr. Pramit Pandit [aut, cre], Ms. Moumita Paul [aut], Dr. Bikramjeet Ghose [aut] |
| Maintainer: | Dr. Pramit Pandit <pramitpandit@gmail.com> |
| Description: | Provides a systematic framework for neural network–based model selection and forecasting using single hidden layer feed-forward networks. It evaluates all possible combinations of predictor variables and hidden layer configurations, selecting the optimal model based on predictive accuracy criteria such as root mean squared error (RMSE) and mean absolute percentage error (MAPE). Predictors are automatically standardized, and model performance is assessed using out-of-sample validation. The package is designed for empirical modelling and forecasting in economics, agriculture, trade, climate, and related applied research domains where nonlinear relationships and robust predictive performance are of primary interest. |
| License: | GPL-3 |
| Encoding: | UTF-8 |
| Imports: | nnet |
| RoxygenNote: | 7.3.3 |
| NeedsCompilation: | no |
| Packaged: | 2026-01-10 06:41:11 UTC; Pramit |
| Repository: | CRAN |
| Date/Publication: | 2026-01-15 17:10:06 UTC |
Neural Network Model Selector
Description
Fits multiple single-hidden-layer neural network models by evaluating all possible predictor combinations and hidden node sizes. The best model is selected based on minimum RMSE on test data.
Usage
nn_model_selector(
data,
response_var,
train_ratio = 0.75,
max_nodes = 10,
maxit = 500,
seed = 123
)
Arguments
data |
A data frame containing the response and predictor variables. |
response_var |
Character string specifying the response variable name. |
train_ratio |
Proportion of data used for training (default = 0.75). |
max_nodes |
Maximum number of hidden layer nodes to evaluate (default = 10). |
maxit |
Maximum number of iterations for neural network training (default = 500). |
seed |
Random seed for reproducibility (default = 123). |
Details
Predictors are standardized before model fitting. Model performance is evaluated using RMSE and MAPE.
Value
A list containing:
best_predictors: Predictor variables of the best model
best_hidden_nodes: Optimal number of hidden nodes
best_performance: RMSE and MAPE of the best model
performance_table: Performance metrics for all model combinations
fitted: Actual vs fitted values for training data
forecast: Actual vs forecasted values for test data
Examples
data_nn <- data.frame(
y = c(
239.7255591, 239.6504622, 239.5848569, 239.5296290,
239.4858835, 239.4547257, 239.4372607, 239.4345936,
239.4478298, 239.4780743, 239.5264322, 239.5940089,
239.6819094, 239.7912389, 239.9231027, 240.0786057,
240.2588534, 240.4649507, 240.6980029, 240.9591152,
241.2493927, 241.5699405, 241.9218640, 242.3062682
),
x1 = c(
9.968768102, 9.160298963, 7.294994564, 5.374395163,
4.640671747, 5.495752064, 7.155488888, 8.532368787,
8.032804811, 10.32506916, 12.17319856, 0.571302071,
12.20714387, 27.13871523, 35.05310057, 42.40476672,
46.28262184, 3.089076495, 40.31650327, 20.83471700,
25.71428597, 21.06398002, 20.26911914, 22.17299909
),
x2 = c(
0.929946922, 4.246863796, 2.895052481, 6.827712819,
11.53788333, 5.688668709, 26.08913871, 30.14926832,
22.77412794, 4.519550904, 18.38195203, 40.50655053,
58.61381025, 69.95404513, 76.08779720, 86.86779542,
79.92326273, 32.26071629, 27.67652481, 66.80672448,
86.54120883, 97.53881465, 95.49058569, 43.06666626
),
x3 = c(
143.7114315, 153.7664088, 158.5007862, 158.7973830,
155.8340003, 150.2453258, 142.4471949, 132.8380705,
121.6890278, 108.8662730, 94.52734991, 78.93448337,
62.31616514, 44.76595425, 26.34367655, 7.109157889,
12.72227903, 32.31332405, 50.67117014, 66.80301029,
79.71603746, 88.41744464, 92.01533759, 90.21350491
)
)
result <- nn_model_selector(
data = data_nn,
response_var = "y",
train_ratio = 0.75,
max_nodes = 5,
seed = 123
)
result$best_performance