Роман Герасимов

Dr. Роман Герасимов

orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
9 September 2025 in Big Data and Cognitive Computing
Verified
Journal Article
Open access
PDF file
EGEkaterina Gribanova
Ekaterina Gribanova
РГРоман Герасимов
Роман Герасимов
EVElena Viktorenko
Elena Viktorenko
This paper introduces novel inverse optimization algorithms (RC and DC) for neural network training in stock price forecasting in an attempt to overcome the traditional gradient descent limitation of local minima convergence. The key novelty is a stochastic algorithm for inverse problems adapted to neural network training, where target function values decrease iteratively through selective weight modification. Experimental analysis used closing price data from 40 Russian companies, comparing traditional activation functions (linear, sigmoid, tanh) with specialized functions (sincos, cloglogm, mish) across perceptrons and single-hidden-layer networks. Key findings show the superiority of the DC method for single-layer networks, while RC proves most effective for hidden-layer networks. The linear activation function with the RC algorithm delivered optimal results in most experiments, challenging conventional nonlinear activation preferences. The optimal architecture, namely, a single hidden layer with two neurons, achieved the best prediction accuracy in 70% of cases. The research confirms that inverse optimization algorithms can provide higher training efficiency than classical gradient methods, offering practical improvements for financial forecasting.

1 October 2025 in IAES International Journal of Artificial Intelligence (IJ-AI)
Verified
Journal Article
Open access
PDF file
EGEkaterina Gribanova
Ekaterina Gribanova
РГРоман Герасимов
Роман Герасимов
This paper presents a new algorithm for building neural network models that automatically selects the most important features and parameters while improving prediction accuracy. Traditional neural networks often use all available input parameters, leading to complex models that are slow to train and prone to overfitting. The proposed algorithm addresses this challenge by automatically identifying and retaining only the most significant parameters during training, resulting in simpler, faster, and more accurate models. We demonstrate the practical benefits of the proposed algorithm through two real-world applications: stock market forecasting using the Wilshire index and business profitability prediction based on company financial data. The results show significant improvements over conventional methods: models use fewer parameters–creating simpler, more interpretable solutions–achieve better prediction accuracy, and require less training time. These advantages make the algorithm particularly valuable for business applications where model simplicity, speed, and accuracy are crucial. The method is especially beneficial for organizations with limited computational resources or that require fast model deployment. By automatically selecting the most relevant features, it reduces the need for manual feature engineering and helps practitioners build more efficient predictive models without requiring deep technical expertise in neural network optimization.

15 January 2026 in Journal of Corporate Finance Research / Корпоративные Финансы
Verified
Journal Article
Open access
PDF file
EGEkaterina Gribanova
Ekaterina Gribanova
EVElena Viktorenko
Elena Viktorenko
РГРоман Герасимов
Роман Герасимов
In the context of modern economic challenges, the effective functioning of the stock market becomes crucial for the sustainable development of the Russian economy. Attracting investments to the real sector requires creating effective tools for market information analysis, which involves processing large volumes of heterogeneous data under conditions of high volatility and geopolitical instability. The aim of the research is to develop an algorithm for building a sparse neural network. This model automatically eliminates insignificant connections between neurons for predicting stock market dynamics. The proposed approachis based on the method of solving a single-point inverse problem with a minimization of the sum of absolute parameter values, which allows reducing the model’s dimensionality. The scientific novelty of the research comprises two aspects. First, the work explores the possibility of using new factors generated by a large language model (an artificial intelligence system for text processing) for predicting stock market dynamics. Second, an original algorithm for constructing a sparse neural network has been developed. The research tested two main hypotheses. The first hypothesis aimed to verify the advantages of sparse neural networks over fully connected architectures in prediction accuracy. The second hypothesis investigated the effectiveness of using features extracted by large language models from unstructured text sources for financial forecasting.Experimental verification on three tasks of stock price and dividend forecasting confirmed both hypotheses. The sparse architecture demonstrated an advantage over fully connected models in prediction accuracy and computational efficiency. Automatic feature selection revealed the relevance of macroeconomic characteristics extracted by the large language model, confirming the promise of integrating modern natural language processing technologies into financial forecasting. The obtained results can be used to form effective strategies of stock market behavior and create intelligent decision support systems. In addition, sparse models can be used in solving other economic problems, including portfolio optimization andfinancial performance management.

21 August 2026 in Intelligenza Artificiale
Verified
Journal Article
Closed access
EGEkaterina Gribanova
Ekaterina Gribanova
РГРоман Герасимов
Роман Герасимов
As data accumulation continues to expand and information technologies evolve, machine learning methods have become widely adopted, making the effectiveness of learning algorithms crucial. Among the most popular machine learning models is Lasso regression, renowned for its feature selection capabilities and ability to address multicollinearity. This paper introduces novel algorithms for estimating Lasso regression parameters by reformulating the problem as an inverse single-point optimization task. Two algorithms are proposed: Lasso-I, which implements coordinate descent with L1 regularization, and Lasso-H, a hybrid approach that combines Lasso-I with wrapper techniques for feature selection using information criteria. The iterative algorithms involve calculating partial derivatives and selecting arguments for adjustment based on residual sum of squares or information criteria. Algorithm evaluation was performed using linear and logistic regression models across diverse datasets from KEEL and UCI repositories, alongside various metrics including the AIC, MSE, and R 2 . The experimental results demonstrate that the algorithms effectively address parameter estimation problems, with Lasso-H achieving optimal AIC values in 90% of logistic regression cases. The proposed methods eliminate the need for explicit regularization parameter specification while maintaining robust feature selection capabilities and effective multicollinearity mitigation, demonstrating high accuracy and reliability across high-dimensional datasets.