Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

Railways Tender Price Prediction Using Machine

Learning and Deep Learning Algorithms
E Pavan Kumar1; Kuzhalini Sivanandam2; Akshay Acharya3; Sandeep Kumar Giri4; Bharani Kumar Depuru5
1,2,3
Research Associate, 4Team Leader, Research and Development, 5Director
Aispry, Hyderabad, India

Abstract:- The idea of determining the optimum bid with attributes. These attributes include tender-specific
amount in any railway tender competition has been a information such as 'Tender_no', 'Railway_Location', 'Zone',
complex and critical task. Reasonable pricing is the main 'Division', 'DateofOpening', 'Nature', 'Awarded_to', and
gateway in the process of winning contracts. The 'L1_Price'. Additionally, the dataset contains detailed pricing
approach presented here is based on data-driven pricing information like 'Basic', 'Qty', 'Category', and various
to maximize tender wins and reduce potential profit loss. customer-related financial metrics ('Customer_1' to
The model shall make use of historical tender data, 'Customer_30').
competitor pricing data, and market indicators in order
to predict the optimal bid amounts. This advanced ML The data analysis revealed that all awarded tenders
model runs advanced machine learning algorithms, resulted in a margin gain of up to 18% to 25% from the basic
including regression models and ensembles, to study the price, with no missing or duplicate values. Among the 30
intricate relationship between factors that would affect customers, only 15 secured tenders, highlighting the
the successful execution of a bid. Deep learning models competitive nature of the bidding process.[1]
are integrated into the model to provide it with better
handling of temporal dependencies and other hidden To ensure the data's integrity and prepare it for
patterns in data, hence yielding accurate and robust modeling, we performed general preprocessing steps,
predictions. including handling null values and removing duplicates[4].
We conducted a thorough analysis to identify the most
The key objective of the work is to increase the influential columns affecting the output, focusing on the
winning rate of tender contracts by at least 10%, while relationships between tender-specific attributes and the final
competitive profitability is ensured. Because of precise awarded price. The preprocessing involved converting
competitor price predictions, business success criteria are 'DateofOpening' to a suitable date format, normalizing
oriented to reaching this higher win rate on tenders. The numerical features such as 'L1_Price', 'Basic', and 'Qty', and
Machine learning success criteria target a price encoding categorical variables like 'Railway_Location',
prediction accuracy of at least 90%. Another important 'Zone', 'Division', and 'Category'. This prepared the dataset
set of economic success criteria that should be targeted is for machine learning algorithms, enabling us to build robust
an improvement in profit margins of at least 5% through models for tender price prediction.
more accurate pricing strategies and a reduction in the
number of rejected bids. The architecture of the project is designed to ensure
scalability and efficiency. Initially, the raw data is stored in a
It will offer great value to the businesses operating in structured database. The preprocessing pipeline will
the railway industry in making proper decisions on preprocess the data and transform this raw data into a suitable
operations and strategic planning. The paper develops a format for initial analysis. Feature selection techniques are
fusion of traditional statistical methodologies with applied to identify the most significant predictors of tender
advanced ML and DL techniques in order to provide a success. The processed data is then fed into various machine
robust solution for competitive advantage and increased learning models [2], including linear regression, decision
profitability in the dynamic and competitive railway trees, and deep learning algorithms like Feed Forward neural
tender market. networks (FFNNs). These models are trained, tested and
validated to predict the optimal bid amounts. Finally, a
Keywords:- Bidding, Tender, Ridge Regression, Feed comprehensive evaluation framework assesses model
Forward Neural Networks. performance based on accuracy, precision, and economic
impact, ensuring the selected model meets the predefined
I. INTRODUCTION business and machine learning success criteria.

Accurate pricing in railway tenders is vital for By employing advanced machine learning techniques,
purchasing contracts while ensuring profitability. In this including regression models and deep learning algorithms,
study, we propose a data-driven approach to predict optimal we aim to accurately predict the optimal bid amounts. This
bid amounts for railway tenders[02,03]. Our dataset pathway doesn’t just increase chances of winning contracts
comprises 20072 entries, each representing a distinct tender from the government but also protects business revenues

IJISRT24JUL1725 www.ijisrt.com 3103

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

because it ensures that costs are kept within reasonable limits. same time contributing towards high quality decision making
The results gained in this way can be very useful when as well as technology- driven operations among other things
evaluating tenders strengths and weaknesses, while at the needed by firms operating railways.

 CRISP-ML(Q): In this Research Article we mainly followed this CRISP-ML(Q) Framework

Fig 1: Project Management Methodology

II. PHILOSOPHY AND METHODS This approach upgraded the solidness and precision of
our conjectures by limiting commotion and zeroing in on
A. Information Collection: huge examples. For information planning, a few
A CSV file containing comprehensive records of preprocessing steps were done to address missing qualities
railway tenders served as the source for the dataset used in and normalize the size of highlights.
this study. The primary target variable in this dataset is
"L1_Price," which represents the lowest bid price for each A pipeline was used to process numerical variables like
tender. Other attributes in this dataset include "Basic" and "Qty," which included scaling with RobustScaler
"Railway_Location," "Zone," "Nature," "Basic," "Qty," and to effectively manage outliers and median imputation to deal
"Category. with missing values. One-hot encoding was used to deal with
categorical variables like "Railway_Location," "Zone,"
B. Transformation and Cleaning of the Data: "Nature," and "Category."
Consider a situation where delicate cost reports are
dispersed with fluctuating opening dates, making it trying to The first step was to substitute the most common value
recognize patterns. To address this, we gathered the data into for missing values. By creating binary columns for each
a monthly format using resampling methods. We were able to category, this encoding technique made categorical data
calculate the average monthly tender prices thanks to this compatible with machine learning algorithms.
change, which gave our predictive models a clearer picture.
Resampling the data to a quarterly format helped reduce These preprocessing steps were applied to the various
short-term fluctuations and highlight long-term trends when feature types in a systematic manner using a Column
daily data became overwhelming. Transformer, resulting in a consistent data transformation
process throughout the various stages of model development.

IJISRT24JUL1725 www.ijisrt.com 3104

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

Imagine you have a pile of tender price documents all promotes model stability and generalizability, making it a
with different opening dates this jumbled mess makes it hard strong choice for datasets with these characteristics.[9, 16]
to see any trends but fear not resampling comes to the rescue
by resampling this data to a monthly format we can combine  Elastic Net: Striking A Balance
the prices into a single clear picture this lets us calculate an The Elastic Net combines both L1 (Lasso) and L2
average monthly tender price which becomes the foundation regularization, offering a powerful tool when feature
for our prediction models these models like detectives selection and coefficient shrinkage are desirable. It leverages
looking for clues use historical patterns to forecast future the strengths of both Lasso and Ridge Regression, often
tender prices here's another scenario you have a mountain of performing exceptionally well on datasets with many features
daily tender prices but you are really interested in the bigger and potential correlations. This makes it a compelling option
picture quarterly trends resampling like a skilled editor can for tender price prediction with a vast array of influencing
help by downsampling the data to a quarterly format we can factors.[16]
condense the daily details and highlight the long-term trends
this smooths out the daily ups and downs giving our  Decision Trees: Unveiling the Reasoning
prediction models a clearer view with less noise to worry Decision Trees model data using a tree-like structure
about the models become more stable and precise leading to where decisions based on features lead to specific outcomes
more reliable forecasts.[6] (tender price range). This approach allows for capturing non-
linear relationships and interactions between features.
To prepare the data for modeling, preprocessing steps Additionally, their interpretability makes it easy for
were implemented to handle missing values and standardize stakeholders to understand the logic behind the predictions.
feature scales. Numerical features, namely 'Basic' and 'Qty', However, they can be susceptible to overfitting on tender
were processed using a pipeline that included median price data with specific nuances.
imputation for missing values followed by scaling with the
RobustScaler to manage outliers effectively.  Random Forest: Ensemble Strength For Enhanced
Accuracy
Categorical features, including 'Railway_Location', Random Forest builds on the foundation of Decision
'Zone', 'Nature', and 'Category', were handled by imputing Trees by creating an ensemble – a group of multiple trees.
missing values using the most frequent value and then Each tree is trained on a random subset of data, promoting
applying one-hot encoding. This encoding technique diversity. By averaging the predictions from these diverse
transformed categorical variables into a format suitable for trees, Random Forest reduces overfitting and often achieves
machine learning models by creating binary columns for each superior accuracy compared to individual Decision Trees. It's
category. robust to noise in the data and adept at handling large,
complex datasets often encountered in tender price
A Column Transformer was used to apply these prediction.[11,12]
preprocessing steps to the respective feature types, resulting
in a unified pipeline that ensures consistent data  Xgboost: Speed and Performance Champion
transformation across different stages. The king of speed and power xgboost is the most
effective and versatile slope helping execution it is
C. Model Development additionally among the most remarkable both concerning
execution and speed it frequently begins at the very beginning
 Model Selection of complex problems as has frequently been observed in
Choosing a Model Picking the right device for the gig tender price predictions gradient boosting involves
is essential in delicate cost expectation. developing models in a logical order with the subsequent
models correcting previous errors because of the way that it
 Linear Regression: The Baseline Benchmark gains from the slip-ups of the past xgboost turns out to be
Linear Regression establishes a straightforward linear truly adept at adapting to mind boggling examples and
relationship between features (project characteristics) and the collaborations existing inside delicate cost information[9,11].
target variable (tender price). It offers a clear and
interpretable model, making it easy to understand how  Gradient Boosting Machines (Gbm): Unveiling
adjustments in features influence the price. While seemingly Complexities
simple, it serves as a valuable baseline model for comparison Gradient boosting machines have evolved as one of the
and can be a good starting point for more complex most potent tools for predicting tender prices across
models.[9,11,12,13] industries the models have an excellent capability for
working over complex data sets and a large number of
 Ridge Regression: Combating Multi Collinearity variables so they are applicable to complex issues like tender
Ridge Regression builds upon Linear Regression by price prediction by improving weak learners sequentially
adding a penalty term (L2 regularization) that discourages normally decision trees gbm manages to capture subtle
overly large coefficients. This is particularly beneficial in relationships between input variables and tender prices it is
situations with high-dimensional data, where features are due to this ability of the model to learn from its past mistakes
highly correlated (multicollinearity). Ridge Regression and focus on the most influential predictors that gbms attain

IJISRT24JUL1725 www.ijisrt.com 3105

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

high predictive accuracy further they deal with numerical and where conventional models may falter when attempting to
categorical data making them accommodate most kinds of understand complex connections within tender price data
tender-related information in diverse aspects thus the however the training process for these networks requires
flexibility and robustness of gbms are most suitably adapted considerable computational power by analyzing complex
to changes in the market and tender specifications providing patterns in tender information artificial neural networks anns
reliable estimates that aid strategic decision-making what can effectively predict tender prices this functionality aids in
really makes gradient boosting machines unique is the making informed decisions throughout the bidding process
accurate prescriptive predictions they are capable of enhancing your likelihood of success while keeping costs in
delivering in an environment where businesses are ever more check anns are designed to manage extensive datasets with
reliant on data-driven insights to negotiate a competitive multiple variables which makes them particularly suitable for
tender process.[5, 9, 14] sectors that handle large volumes of tender-related data think
of it as employing a highly capable computer to sift through
 Support Vector Regression (SVR): Dealing With Outliers all that information and deliver the insights necessary for
And Non-Linearities improved forecasting once developed anns can streamline the
Of late support vector machines have proved to be quite tender price prediction process allowing your team to
a strong tool in the prediction of tender prices within the conserve time and reduce the risk of errors picture the
construction environment these models function effectively increased efficiency in fast-moving industries where timely
on data-sets that are complex high dimensional very common decision-making is crucial while cultivating anns requires
when estimating project costs in essence svms work on the specialized knowledge and significant processing resources
basis of knowing a best plane that separates different points the advantages they offer far surpass the initial costs consider
or in this case using previous data and given variables such as it a strategic investment in your organizations long-term
project size location materials and labor cost to predict the success in todays competitive landscape utilizing a robust
tender price thus it can offer its user accurate price estimates tool like anns can significantly impact your ability to secure
by capturing the underlying relationships between these important tenders therefore recognizing the promise of these
variables thereby enabling appropriate budgeting and neural networks is essential they could be the advantage you
selection of competitive bids for contractors and project need to remain ahead of the competition.[8, 9, 10, 12]
managers it has been shown that svms work very well with
outliers and provide accommodation to nonlinear  Long Short-Term Memory: Analyzing Sequential Data
relationships through the use of kernel functions hence Conventional methods for tender pricing often fail to
finding versatility in various tender prediction scenarios fully represent the complexity involved this where LSTM
needless to say generalizing to new data and performing networks become relevant these advanced neural networks
robustly in high-dimensional spaces svms stand out in the are proficient at analysing continuous data enabling them to
competitive landscape of predictive modeling techniques in reveal complex designs as well as,relationships within
tender pricing for construction projects.[11,12] previous tender rate of the facts, this cruciality is essential
since cost archive trends can have a substantial crash on the
 Feedforward Neural Networks: Untangling Complexities outcome of fortune bids by incorporating before collections
Feed Forward neural networks are made up of layers of of data LSTM can consider market fluctuations and
interconnected processing units nodes that learn intricate competitor behaviour producing a deeper insight into the
patterns and relationships among features through non-linear bidding environment.[7,15]
transformations this capability enables them to excel in tasks

 Architecture Diagram

Fig 2: Architecture Diagram

IJISRT24JUL1725 www.ijisrt.com 3106

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

A Feedforward Neural Network (FFNN) was selected E. Model Training and Evaluation
due to its capacity to capture complex patterns and
interactions within the data. The FFNN's architecture was  Early Stopping
designed to address the non-linearity of the prediction Early Ending in Sensitive Worth Assumptions To work
problem, making it a suitable choice for predicting tender on the show and capability of our fragile expense assumption
prices. models, we executed an early stopping instrument during the
readiness stage. Overfitting is a typical issue where a model
 Pipeline Creation loses its capacity to sum up to new, concealed information
The preprocessing and modeling steps were when it turns out to be excessively custom fitted to the
encapsulated in a pipeline to streamline the workflow from preparation information. This prevents this from happening.
data preparation to model training. This pipeline was saved The idea of early stopping: When the model's presentation on
as a pickle file to facilitate reproducibility and ease of an approval set stops progressing to the next level, early
deployment. stopping is meant to end the preparation cycle. Essentially,
during setting up, the model is evaluated on an alternate
 Train-Test Split endorsement dataset at standard stretches. After a
The dataset was partitioned into training and testing sets predetermined number of epochs (training iterations), the
using an 80-20 split. The training set was employed for model training is ended if the validation loss—an indicator of the
development and hyperparameter tuning, while the testing set model's performance on this unseen data—does not
was reserved for assessing the model's performance on significantly improve. With this approach, it is possible to
unseen data. avoid situations in which the model's performance is harmed
by overfitting or continued training only yields modest gains.
D. Hyper Parameter Tuning
Application in Our Environment: Early stopping was
 Model Architecture meticulously implemented in our project to monitor
The FFNN model's architecture was defined with an validation loss throughout the training process. We spread out
input layer matching the number of transformed features, a resilience limit, which describes the amount of ages the
followed by three hidden layers. The number of units, model can continue planning without seeing an improvement
activation functions, and dropout rates in these layers were in endorsement hardship. If this patience threshold was
treated as hyperparameters for optimization. reached and no further gains were observed, the training was
immediately stopped. The model was prepared to be
 Hyper Parameter Optimization sufficiently lengthy to gain from the information without
Hyperband optimization was employed to fine-tune the squandering assets or risking overfitting thanks to this
model's hyperparameters. This method efficiently searches procedure.
through the hyperparameter space to identify the optimal
configuration by evaluating multiple models with varying  Advantages of Stopping Mid:
hyperparameters over several epochs.
 Prevents Overfitting: By stopping training as soon as the
 Hyperparameter Trials model's performance on the validation set begins to
Several trials were conducted to find the best model plateau or worsen, early stopping helps maintain the
configuration. The following are summaries of some of the model's ability to generalize to new data rather than
trials: becoming overly specific to the training set.
An initial working model specified by these parameters  Saves Resources: Early stopping reduces unnecessary
has been tested: First layer –192 nodes –‘relu’, drop out =0.3; computation by terminating training when additional
Second Layer—320 nodes —‘sigmoid’,drop out=0.2; Third epochs no longer contribute to better performance. While
Layer—416 nodes—‘relu’, drop out=0.2 .The score for this working with enormous datasets or complex models, this
trial was calculated to be equal to11 .220. productivity is particularly helpful.
 Improves Model Robustness:Building a more powerful
In this case I have specified that: First Layer—256 indicator at delicate costs is made easier by ensuring that
Nodes—‘ReLU”, Drop Out=0.4 ; Second Layer—64 Nodes the model is prepared appropriately. It accomplishes a
- “ReLU”, Drop Out=0.3 model with incredible execution across an assortment of
datasets by finding some kind of harmony among
 Cross-Validation underfitting and overfitting. Including early stopping in
To ensure robust evaluation, 10-fold cross-validation our training gave us a strategic advantage by preventing
was employed. The training set was divided into 10 folds, overfitting, maximizing resource utilization, and
with each fold used as a validation set in turn while the increasing the overall reliability of our tender price
remaining folds were used for training. This approach prediction models.
provided a comprehensive evaluation of the model's
performance across different subsets of the data.

IJISRT24JUL1725 www.ijisrt.com 3107

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

An early stopping mechanism was integrated to  R SQUARE SCORE: The R score portrays the proportion
terminate training when the validation loss no longer of variance in the target variable predictable from the
improved over a specified number of epochs. This strategy independent variables an R near 1 shows a model that
prevented overfitting and ensured that the model was trained explains a great deal of variability in the target variable
only as long as necessary. FFNN led at the very best with an R rating thus showing
functionality within its capability to seize and model
 Model Evaluation underlying developments in gentle charges additional.
Our fashions effectivity will probably be assessed to be
efficient with a mix of general efficiency metrics and Evaluation strategies educate-take a look at break up to
evaluation methods to make sure that they meet the outlined ensure that our models generalize nicely to new records we
objectives and ship dependable predictions we zeroed in on a divided the dataset into training and checking out subsets with
few key metrics suggest squared errors suggest absolute an 80-20 split this method allowed us to evaluate expertise
mistakes suggest absolute pc mistakes and r ranking these nicely the models finished on unseen statistics confirming
metrics gave a complete view of every mannequins accuracy their robustness and applicability cross-validation 10-fold
and dependability. skip-validation was conducted to evaluate the stability and
reliability of the fashions by splitting the education statistics
 Mean Squared Errors(MSE) : MSE is the average of the into ten subsets schooling the version on 9 of them and
squares of the errors its the common squared difference validating it at the remaining one we received extra nuanced
between predicted and actual values a low mse means that details of each versions ordinary overall performance
the versions predictions are closer to the actual values throughout one-of-a-kind information subsets early stopping
some of the models tested the feedforward neural an early stopping mechanism was included to prevent
networks confirmed the lowest mse reflecting their high overfitting and ensure maximum first-class education
benchmark performance overall in minimizing mistakes duration this approach stopped education as quickly as the
in predictions. validation loss plateaued thus retaining model generalization
 Mean Absolute Error(MAE): MAE method measures and warding off useless computational fee basic the
typical significance of mistakes in predictions simply not evaluation has highlighted that feedforward neural networks
taking their course into consideration it indicates the did not just excel in terms of accuracy information they also
typical mistake in simple absolute terms the FFNN also tested steady overall performance across unique metrics this
rated finest with this metric as it gave the bottom mae makes them the maximum dependable choice for tender price
which proved its accuracy for predicting light prices. prediction for many of the models examined by focusing on
 Mean Absolute Percentage Error(MAPE): MAPE is a those complete evaluation metrics and methods we ensured
measure of the precision of a model as a percentage that our final model would perform well not only in terms of
capturing how far the load predictions are off from their know-how but also according to the high standards required
actual values compared to their cost. This metric is useful for accurate and efficient smooth price predictions.
for prediction basics.
 Overall performance in percentage phrases FFNN turned The best model identified through hyperparameter
out with the smallest MAPE which indicates that it is tuning was evaluated on both the training and testing sets.
effective in making correct percentage-based total Key performance metrics included Mean Squared Error
predictions. (MSE), Mean Absolute Error (MAE), Mean Absolute
Percentage Error (MAPE), and R2 score. The following
metrics were reported:These metrics demonstrate the model's
strong performance and its ability to generalize well to new
data.

Fig 3: Model Metrics

IJISRT24JUL1725 www.ijisrt.com 3108

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

F. Model Deployment In this research on estimating tender prices, we

For deploying the Feedforward Neural Network considered various machine learning models. Linear
(FFNN) model using Streamlit, a comprehensive approach Regression had moderate performance, indicating
was implemented to ensure an efficient and user-friendly underfitting with training and test accuracies of 79% and
interface. The model, chosen for its proficiency in capturing 65%, respectively. Ridge Regression improved
complex patterns within the tender price data, was developed generalization, achieving 88% and 80% accuracy. Elastic Net
and fine-tuned through meticulous steps. A well-structured struggled with the bias-variance trade-off, resulting in lower
pipeline encapsulating preprocessing and modeling steps was performance. The Decision Tree Regressor, while highly
created, ensuring reproducibility and ease of deployment. The accurate on training data (92%), showed signs of overfitting
dataset underwent an 80-20 train-test split to facilitate model with a test accuracy of 81%. Random Forest Regressor
training and performance assessment. Hyperparameter tuning performed robustly with high accuracies of 93% (train) and
was conducted using Hyperband optimization, which 85% (test) due to its ensemble approach. XGBoost and
efficiently searched through the hyperparameter space to Gradient Boosting Machines (GBM) demonstrated good
identify the optimal model configuration. Several trials with results, although XGBoost required careful tuning to avoid
varying units, activation functions, and dropout rates were overfitting. Support Vector Regression (SVR) captured non-
evaluated, ensuring the selection of the best performing linear relationships well but showed overfitting, with a 13%
model. A robust evaluation was performed using 10-fold gap between training (89%) and test (76%) accuracy.
cross-validation to provide a comprehensive assessment of Feedforward Neural Networks excelled, achieving the
the model's performance across different data subsets. Early highest accuracies of 94% (train) and 91% (test) due to their
stopping was integrated to prevent overfitting, ensuring the ability to model complex patterns. Overall, balancing model
model was trained only as long as necessary. Key complexity and generalization is crucial for optimal
performance metrics such as Mean Squared Error (MSE), performance in tender price prediction.
Mean Absolute Error (MAE), Mean Absolute Percentage
Error (MAPE), and R2 score were reported to demonstrate IV. CONCLUSION
the model's efficacy. The final step involved deploying the
model using Streamlit, providing an interactive platform for Hence this review streamlines the use of cutting edge
users to input project characteristics and obtain tender price ML and DL calculations to upgrade the precision and
predictions seamlessly. This deployment not only enhanced viability of delicate cost assessment in the rail line area
accessibility but also ensured that the model's sophisticated coordinating a different arrangement of models comprising of
capabilities could be utilized effectively by stakeholders in direct relapse edge relapse choice trees irregular woodlands
the tendering process. xgboost slope helping machines support-vector-relapse and
feedforward-brain network models we distinguished the most
III. RESULTS AND DISCUSSION suitable methodologies for anticipating ideal bid sums our
outcomes demonstrate that while less complex models
In this quest to evaluate sensitive charges several ai including straight and edge relapse give a standard more
models were tried direct regression had quite a decent modern procedures like irregular timberlands xgboost and
performanceit underfitted training and testing accuracies feedforward-brain networks calculations further develop the
were 79% and 65% respectively ridge regression performed expectation exactness essentially among these the best-
well given that it yielded an accuracy of 88% and performing model has been feedforward brain networks with
generalization rates of 80% elastic net struggled with the bias- the most noteworthy precision rates for both the preparation
variance trade-off hence performing poorly while the choice and testing datasets our outcomes mirror that this class of
tree regressor was extremely accurate on training data to the models gives the vital model ability to catch complex
tune of 92% preparatory information it gave indications of examples and cooperations in information that are urgent for
overfitting with a test accuracy of 81% random forest making precise and cutthroat bid offers the class of models
regressor did well due to its ensemble approach train utilized in this application gives a model limit that will
accuracies were 93% while test accuracies came in at 85% actually want to catch complex examples and connections in
xgboost and gradient boosting machines turned in good the information that are important to offer precisely and
results but care was required in tuning xgboost to avoid seriously utilization of this class of models further profits a
overfitting support vector regression did a fine job capturing hearty structure in foreseeing the delicate costs as well as in
the nonlinear relationships but suffered from overfitting as offering decisively helping the organizations endeavors to
shown by the 13% gap between train accuracy of 89% and ideally situate itself in aggressive offering circumstances as
test accuracy of 76% feedforward neural networks performed such we recommend the utilization of information driven bits
very well due to their inherent modeling of complex of knowledge in blend with cutting edge ml and dl methods
relationships achieving the highest accuracy of 94% on the to enhance the offering methodology for rail route firms
train set and 91% on the test set in making any tender price thusly profiting from an expanded achievement rate with very
predictions balance needs to be maintained between model high bid achievement rates and limiting bid dismissal to
complexity and generalizability. guarantee serious productivity obviously this present reality
application to support future enhancements and versatility to
showcase changes in delicate details ought to be borne as a
primary concern.

IJISRT24JUL1725 www.ijisrt.com 3109

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL1725

In this research work a feed-forward neural network [17]. Model selection Feed Forward Neural Networks for
FFNN model was developed and validated to predict railway Forecasting inflow and outflow in Indonesia.
project tender prices optimization of the model used Suhartono, Prilyandari Dina Saputri, Farah Fajrina
hyperband while cross-validation was employed for a proper Amalia, Dedy Dwi Prastyo and Brodjol Sutijo Suprih
evaluation the strategy ensures robust execution and is able Ulama.
to offer a reasonable device to delicate cost expectation hence https://link.springer.com/chapter/10.1007/978-981-
showing practicality for cutting-edge machine learning 10-7242-0_8
methods within this space. [18]. Effective House Price Predictions using Machine
Learning.Jincheng Zhou, Tao Hai, Ezinne
REFERENCES C.Maxwell-Chigozie, Afolake Adedayo, Ying Chen,
Celestine Iwendi and Zakaria Boulouard.
[1]. Predicting the Tender Price of Buildings during Early https://link.springer.com/chapter/10.1007/978-3-031-
Design: Method and Validation. R. McCAFFER, M. 37164-6_32
J. McCAFFREY and A. THORPE [19]. Comparative Analysis of Random Forest Regression
[2]. https://www.jstor.org/stable/2581370?seq=1&cid=pd for House Price Prediction. Obilisetti Lohith,Aman
f-reference#references_tab_contents Jha and Shamstabrej Chand Tamboli
[3]. Multiple Linear regression model for predicting https://ijcrt.org/papers/IJCRT2306866.pdf
bidding price Petrovski Aleksandar, Petruseva [20]. Multiple Linear Regression Model for Prediction
Silvana, Zileska Pancovska Valentina Bidding Price. Petrovski Aleksandar,Petruseva
https://www.researchgate.net/publication/282646773 Silvana and Zileska Pancovska Valentina
_Multiple_Linear_regression_model_for_predicting_ https://www.researchgate.net/publication/282646773
bidding_price _Multiple_Linear_regression_model_for_predicting_
[4]. Predicting Bidding Price in Construction using bidding_price
Support Vector Machine [21]. Gradient Boosting Censored Regression for Winning
[5]. Silvana Petruseva, Phil Sherrod, Valentina Zileska Price Prediction in Real-time Bidding . Piyush Paliwal
Pancovska, Aleksandar Petrovski and Olesksii Renov
[6]. https://dx.doi.org/10.18421/TEM52-04k https://link.springer.com/chapter/10.1007/978-3-030-
[7]. A machine learning-based Bidding price optimization 18590-9_43
algorithm approach Saleem Ahmad, Sultan Salem, [22]. A Machine Learning Based Bidding price
Yousaf Ali Khan, I.M. Ashraf optimization algorithm approach. Saleem Ahmed,
[8]. https://doi.org/10.1016/j.heliyon.2023.e20583 Sultan Saleem, Yousaf Ali Khan and I.M.Ashraf
[9]. Gradient Boosting Censored Regression for Winning https://www.cell.com/heliyon/fulltext/S2405-
Price Prediction in the Real-Time-Bidding, Piyush 8440(23)07791-
Paliwal and Oleksii Renov 5?_returnURL=https%3A%2F%2Ffanyv88.com%3A443%2Fhttps%2Flinkinghub.elsevi
[10]. https://link.springer.com/chapter/10.1007/978-3-030- er.com%2Fretrieve%2Fpii%2FS2405844023077915
18590-9_43 %3Fshowall%3Dtrue
[11]. Resampling strategies for imbalance time series [23]. Regularization and variable selection via the elastic
forecasting. Nuno Moniz, Paula Branco, Luís Torgo net. Hui Zou, Trevor Hastie
https://link.springer.com/article/10.1007/s41060-017- https://rss.onlinelibrary.wiley.com/doi/10.1111/j.146
0044-3 7-9868.2005.00503.x
[12]. Revolutionizing TMT Steel Bar Sales Projections:
Unleashing the Power of Deep Learning Algorithms
for Unparalleled Forecasting Precision.
[13]. Mohd Amer Hussain, Akhil Rasamsetti, Vaanishree
Kamthane, Deba Chandan Mohanty
[14]. https://ijisrt.com/assets/upload/files/IJISRT23NOV21
92.pdf
[15]. Market Clearing Price Prediction Using ANN in
Indian Electricity Markets. Anamika & kumar
https://sci-
hub.se/https:/doi.org/10.1109/ICEETS.2016.7583797
[16]. Advanced Machine Learning Algorithms for House
Price Prediction: Case Study in Kuala Lumpur.
Shuzlina Abdul-Rahman, Sofianita Mutalib,Nor
Hamizah Zulkifley and Ismail Ibrahim
https://thesai.org/Downloads/Volume12No12/Paper_
91-Advanced_Machine_Learning_Algorithms.pdf

IJISRT24JUL1725 www.ijisrt.com 3110