Quant ClubQuant Club
BlogsWhitePaperAbout UsContact
SoQ

Search

Search for a command to run...

Quant Research

Research Papers &
Whitepapers

Academic research on quantitative finance, algorithmic trading, and portfolio construction by the researchers at Quant Club IIT Kharagpur.

View Blog

Latest Paper

Optimization of portfolios with cryptocurrencies: Markowitz and GARCH-Copula model approach
FeaturedResearch Paper

Optimization of portfolios with cryptocurrencies: Markowitz and GARCH-Copula model approach

The rapid growth and increasing popularity of cryptocurrencies have sparked considerable interest in their potential role within traditional investment portfolios. As cryptocurrencies are characterized by high volatility and distinct risk-return profiles, traditional portfolio optimization methods, such as the Markowitz Mean-Variance framework, may struggle to provide accurate risk assessments and optimal allocation strategies. The Markowitz model, while foundational in portfolio theory, assumes normally distributed returns, a characteristic often not observed in cryptocurrencies. Meanwhile, the GARCH-Copula model, with its ability to model time-varying volatility and capture complex dependencies between assets, provides a more flexible approach. By combining these models, an enhanced portfolio optimization framework can be developed that better accommodates the idiosyncratic behavior of cryptocurrencies, potentially offering improved portfolio performance and risk management for investors. In this work, GARCH-Copula (vine copula) was combined with Markowitz optimization. Three types of portfolios were optimized: a traditional portfolio containing stocks, gold, and commodities; a cryptocurrency portfolio; and a combined portfolio of both traditional and cryptocurrency assets. Markowitz optimization was applied with the aim of maximizing the Sharpe ratio, enhancing the risk-return tradeoff for each portfolio.

June 2, 2026

More Papers

Stock Price Prediction Using Long Short-Term Memory Graph Convolutional Neural Networks (LSTM-GCNs)
Research Paper

Stock Price Prediction Using Long Short-Term Memory Graph Convolutional Neural Networks (LSTM-GCNs)

Stock price prediction is a fundamental aspect of financial trading, enabling traders to make informed decisions about buying, selling, and holding stocks. Traditional models often fail to account for the intricate relationships between stocks, such as sectoral correlations or value- chain dependencies. In this paper, we propose a novel neural network-based model, the Long Short-Term Memory Graph Convolutional Neural Network (LSTM-GCN), which combines the temporal modeling capabilities of LSTM networks with the spatial modeling capabilities of Graph Convolutional Networks (GCNs). The LSTM-GCN model leverages sectoral and value-chain relationships to predict stock returns more accurately. We validate our approach using historical data from two major stock markets, NYSE and NASDAQ, demonstrating its superior performance in capturing complex temporal and spatial dependencies compared to traditional methods.

June 2, 2026
Generating volatility surfaces using Variational Autoencoders
Research Paper

Generating volatility surfaces using Variational Autoencoders

Volatility of an asset is a metric of immense importance in financial markets, serving as a key indicator of risk and uncertainty. It is widely utilized by a broad spectrum of market participants, including traders, investors, portfolio managers, and risk analysts for decision-making in the financial markets. A common and convenient way in which volatility is visualized and used is in the form of Volatility surfaces. It is dependent on Implied Volatility of the asset’s options. A volatility surface shows the volatility that is implied by the market price of an option on an asset as a function of the options strike price and maturity. However, constructing a reliable volatility surface is often hindered by missing or noisy data, which can distort insights and affect decision-making. To resolve this problem in the case of a volatility surface, we can use Variational Autoencoders to estimate unobservable or missing parts of the surface. By learning latent representations of the data, VAEs can capture the underlying structure of the volatility surface and accurately fill in missing values.

June 2, 2026
Deep BSDE Solvers and High-Dimensional PDEs
Research Paper

Deep BSDE Solvers and High-Dimensional PDEs

This white paper provides an in-depth exploration of two seminal research papers: E, Han, and Jentzen (2017): Deep BSDE solvers for high-dimensional parabolic partial differential equations Han, Jentzen, and E (2018): Solving high-dimensional partial differential equations using deep learning Building on these works, I have elaborated on deep learning-based numerical methods for solving high-dimensional parabolic PDEs via their backward stochastic differential equation (BSDE) formulations. I have retained mathematical rigor while providing clear, step-by-step explanations of the methodology. Further, it is discussed how these methods apply to quantitative finance, illustrate practical considerations with mathematical insights, and highlight future directions. I have also detailed the improvements and different approaches taken compared to the original papers, incorporating additional knowledge from the field to enhance understanding and potential advancements.

June 2, 2026
From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection
Research Paper

From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection

Time series forecasting is crucial for decision-making across various domains, but traditional methods struggle with disruptions caused by external events and fail to connect social dynamics with data fluctuations. Integrating news articles into forecasting enriches predictions with real-world context, capturing qualitative influences like policy changes and sentiment shifts. This paper proposes a unified approach that embeds news and time series data using fine-tuned large language models (LLMs). By combining quantitative and qualitative insights, the model adapts to dynamic environments, improving accuracy and reliability. The approach emphasizes effective news filtering and reasoning, leveraging LLM agents to analyze, select, and integrate relevant news dynamically, enabling context-aware forecasting.

June 2, 2026
Enhancing Portfolio Allocation by LSTM Model
Research Paper

Enhancing Portfolio Allocation by LSTM Model

Accurate stock price forecasting and portfolio optimization are critical for investors aiming to maximize returns while managing risk. This whitepaper presents a comprehensive framework that integrates Long Short-Term Memory (LSTM) neural networks with Monte Carlo simulations, Efficient Frontier analysis, and key risk-return metrics to achieve optimal investment strategies. The LSTM model is employed to predict future stock prices by leveraging its ability to capture complex temporal patterns in historical price data. Using the forecasted prices as input, Monte Carlo simulations generate a wide range of potential portfolio outcomes, providing insights into performance under various market scenarios. Efficient Frontier analysis is then applied to identify portfolios that offer the best risk-return trade-offs, specifically focusing on the Minimum Variance and Maximum Sharpe Ratio portfolios. Key risk-return metrics, including expected returns, standard deviation, and Sharpe Ratio, are used to evaluate the optimized portfolios. The methodology is demonstrated using stocks from leading technology companies like Apple Inc. (AAPL), Meta Platforms Inc. (META), Amazon.com Inc. (AMZN), Microsoft Corporation (MSFT), and NVIDIA Corporation (NVDA) from January 1st ,2018 to May 31st, 2023, and results indicate that the proposed strategy significantly outperforms S&P 500.

June 1, 2026
Predictive Maintenance with Financial Integration: A Machine Learning Approach
Research Paper

Predictive Maintenance with Financial Integration: A Machine Learning Approach

This paper aims to address the inefficiencies inherent in reactive and preventive maintenance by developing a predictive maintenance strategy that leverages machine learning techniques. It further integrates financial metrics into the analysis, providing a holistic evaluation of how predictive maintenance can optimize both operational efficiency and cost-effectiveness, thus supporting informed decision-making in asset management.

June 1, 2026
A Practical Framework for High-Frequency Market Making: An Adaptation of the Avellaneda & Stoikov Model
Research Paper

A Practical Framework for High-Frequency Market Making: An Adaptation of the Avellaneda & Stoikov Model

This white paper presents a well-known quantitative framework for high-frequency (HFT) market making, inspired by the seminal work of Avellaneda & Stoikov (2008). The model balances expected profit from setting bid-ask quotes against the risk of holding undesired inventory. By framing the market making problem through a continuous-time stochastic control lens, we derive optimal strategies for quoting prices. This paper aims to distill the mathematical model into intuitive concepts, provide guidance on practical implementation, and present a simple coding example to empower practitioners to adapt these strategies to real-world trading environments.

June 1, 2026
Physics-Constrained Deep Learning Framework for Forecasting Wheat Futures Returns
Research Paper

Physics-Constrained Deep Learning Framework for Forecasting Wheat Futures Returns

This project is a four-stage physics driven machine learning pipeline developed to forecast wheat futures returns by integrating satellite remote sensing data, meteorological data and financial data using sophisticated deep learning models. The framework uses more than 40 years of Landsat imagery (1984–2024) together with ERA5-Land climate reanalysis data to capture long-term crop growth behavior and environmental and geological factors. The pipeline converts raw and noisy satellite and Earth observation data into more meaningful data for the signals. First, a Causal Data Engine removes temporal bias by applying strict time-based processing and expanding-window standardization, ensuring no look ahead. Second, a Feature reduction stage reduces noise and dimensionality using the robust denoising autoencoders. Third, a Physics-Constrained Neural Network (PCNN) models crop growth dynamics under biophysical constraints and the governing differential equation. Finally, a commodity returns forecasting module combines the physics-derived state with market data using BiLSTM architecture to estimate future wheat return distributions. By unifying remote sensing, climate physics, and financial time-series modeling, This pipeline provides a robust and interpretable approach for agricultural commodity price forecasting.

June 1, 2026
Queue Reactive Model
Research Paper

Queue Reactive Model

Large order execution in financial markets requires balancing execution certainty against price improvement — a trade-off that static strategies like TWAP and VWAP handle poorly due to their non-adaptive nature. This paper presents the Queue-Reactive Model (QRM), which reframes execution as a queue positioning problem rather than a timing one. By modeling the Limit Order Book as a system of state-dependent stochastic queues and formulating optimal quoting as a Hamilton-Jacobi-Bellman control problem, QRM dynamically adapts to real-time market conditions, explicitly quantifies execution risks, and identifies the optimal spread capture vs. execution speed trade-off. A Reinforcement Learning extension further enables policy learning without explicit model assumptions. QRM is shown to outperform traditional benchmarks while providing theoretical guarantees grounded in market microstructure.

June 1, 2026
Answer-Type-Aware Evaluation of LLM Hallucinations in Financial Question Answering
Research Paper

Answer-Type-Aware Evaluation of LLM Hallucinations in Financial Question Answering

The issue of hallucinations has been recognized as fundamental deficiency in large language models (LLMs). This problem if left unchecked is a huge problem, especially in fields like finance, medicine or law. In this paper, we present a systematic evaluation of multiple LLMs performance on financial question datasets by explicitly separating numerical extraction and reasoning and textual explanation tasks. We evaluate multiple open source LLMs with answer type specific metrics which includes exact-match accuracy and claim-level hallucination rate for all kinds of questions, and numerical and textual questions individually. The LLMs are provided with the context pertaining to the question along with the question and instructed to strictly use only this information. Our results reveal pervasive arithmetic errors in numeric answers and some unsupported claims made in textual explanations. Moreover, no single model performs robustly across both answer types, indicating inherent specialization. These results highlight the limitations of general-purpose LLMs in financial domains and underscore the need for task-aware evaluation and modular model selection in financial QA systems.

June 1, 2026
MGF-LSTM: An Adaptive Hybrid Multi-GARCH Fusion Framework for Volatility Forecasting
Research Paper

MGF-LSTM: An Adaptive Hybrid Multi-GARCH Fusion Framework for Volatility Forecasting

Accurate volatility forecasting is essential for financial risk management, yet traditional econometric models often struggle with the non-linear dynamics of emerging markets. This paper introduces the MGF-LSTM: Multi-GARCH Fusion LSTM, a novel framework that synergizes the statistical precision of GARCH family estimators with the adaptive capabilities of deep learning. Unlike conventional static ensembles, our architecture employs a dynamic attention mechanism to adaptively weigh asymmetric volatility signals (EGARCH, TGARCH) and market sentiment indicators (India VIX) in real-time. By optimizing a robust Quasi-Likelihood (QLIKE) loss function against drift-independent Yang-Zhang volatility proxies, the model captures complex stylized facts such as leverage effects and volatility clustering more effectively than standalone baselines. Empirical evaluation on the NIFTY 50 index confirms the model’s superior performance, achieving a Mean Absolute Percentage Error (MAPE) of 9.19% and a highly calibrated 95% Value-at-Risk breach ratio of 5.73%. Keywords: Volatility Forecasting; GARCH; Attention Mechanism; VaR Analysis.

June 1, 2026
Regime-Augmented Signal Preprocessing for Deep Reinforcement Learning
Research Paper

Regime-Augmented Signal Preprocessing for Deep Reinforcement Learning

Financial time series are simultaneously noisy and non-stationary, defeating naive deep reinforcement learning. Prior research established that wavelet-based signal preprocessing dramatically improves RL learnability, with static Coiflet-4 denoising achieving optimal performance on S&P 500 E-Mini futures. However, this approach assumes markets are stationary—a violated assumption. We extend this work by injecting probabilistic regime awareness directly into the signal preprocessing layer. Rather than switching wavelets via brittle thresholds, we augment the RL observation space from 3 dimensions (denoised signals) to 6 dimensions by appending HMM-inferred regime posteriors. An unsupervised Gaussian HMM learns three market regimes from raw signals; regime posterior probabilities P(zt) are concatenated with wavelet-denoised DIX, GEX, and VIX, producing a regime-contextual observation vector. On 14 years of out-of-sample S&P 500 futures data (709 test days, including COVID-19 crash), regime-augmented preprocessing achieves 94.37% total return and 1.258 Sharpe ratio versus 52.68% return and 0.979 Sharpe for baseline PPO. This 41.69 percentage-point improvement and 0.279 Sharpe gain derive purely from state augmentation, not architectural innovation: the RL algorithm, network architecture, and hyperparameters remain identical across all comparisons. We emphasize that this work extends prior preprocessing research; it does not propose novel RL architectures or regime modeling. The contribution is methodological: demonstrating that probabilistic regime context in the observation space substantially improves learning stability under non-stationarity.

June 1, 2026
GNN for quantitative finance
Research Paper

GNN for quantitative finance

Modern financial markets are inherently networked systems where assets, institutions, and participants are inter-connected through complex relationships. Graph Neural Networks (GNNs) offer a principled approach to modeling these relational structures, capturing phenomena such as volatility spillovers, credit contagion, and momentum effects. This paper surveys GNN methodologies applied to quantitative finance, systematizing graph types, application domains, and empirical results. While GNN-based models demonstrate significant improvements—achieving 70%+ accuracy on long-term stock prediction and 9–23% better performance on volatility forecasting—we identify critical fallbacks including graph construction sensitivity, non-stationarity vulnerability, and scalability constraints. We analyze why GNNs are theoretically suited for financial modeling and critically evaluate the gap between academic benchmarks and realistic trading scenarios. This survey bridges relational reasoning and temporal dynamics in financial markets, identifying open research challenges and practical limitations.

June 1, 2026
Adaptive prompting and trust aware memory
Research Paper

Adaptive prompting and trust aware memory

Prompt engineering has emerged as a powerful paradigm for eliciting complex reasoning behavior from large language models (LLMs) without parameter fine-tuning. However, recent reasoning-oriented prompting strategies such as Chain-of-Thought and Tree-of-Thoughts often incur substantial computational cost and remain vulnerable to error amplification when external memory or retrieval mechanisms are employed. This paper investigates two fundamental questions. First, can adaptive prompting preserve the reasoning capability of advanced prompting strategies while significantly reducing token usage? Second, can trust-aware memory mechanisms reduce error amplification caused by noisy or contradictory retrieved information? To address these questions, we propose two complementary system-level mechanisms: (i) an Adaptive Prompting Controller that dynamically allocates reasoning depth based on estimated uncertainty, and (ii) a Trust-Aware Memory mechanism that filters retrieved context using trust scores and contradiction detection. Through controlled experiments, we demonstrate that adaptive prompting avoids unnecessary reasoning escalation on simpler tasks, yielding substantial cost savings, while trust-aware memory improves robustness under noisy retrieval. Together, these results suggest that scalable and reliable LLM reasoning requires prompt orchestration and memory reliability modeling rather than static prompt design alone.

June 1, 2026
Adaptive Liquid Time-Constant Networks with Dynamic Graph Learning for Multi-Regime Financial Forecasting
Research Paper

Adaptive Liquid Time-Constant Networks with Dynamic Graph Learning for Multi-Regime Financial Forecasting

Financial time series forecasting remains challenging due to regime shifts, non-stationarity, and complex inter-asset dependencies. We present a novel hybrid architecture combining Liquid Time-Constant (LTC) networks, Graph Neural Networks (GNN), and N-HiTS for multi-horizon forecasting across diverse market regimes. Our model achieves a test/train loss ratio of 7.88× on a 10-asset portfolio spanning 2000-2023, demonstrating graceful degradation during unprecedented regime shifts including the COVID-19 crash and 2022 inflation crisis. Key contributions include: temporal regime-aware data splitting that prevents information leakage, dynamic graph structure learning for evolving asset correlations, spectral normalization for Lipschitz-constrained forecasts, and comprehensive regularization preventing catastrophic overfitting. The model exhibits asset-specific performance patterns, excelling at defensive assets (TLT MAE: 0.46, GLD MAE: 0.30) while showing conservative bias on growth equities during bubble periods—a desirable risk management property. We demonstrate 1.57× MAE growth over 400 test samples, significantly outperforming baseline LSTM (25× overfitting ratio) and classical GARCH models.

June 1, 2026
Distributional RL
Research Paper

Distributional RL

Most reinforcement learning (RL) approaches for algorithmic trading optimize for expected returns, effectively ignoring the higher-order moments of the return distribution. In financial markets, however, downside risk and tail events are critical determinants of a strategy’s viability. This white paper presents a risk-sensitive trading agent built upon the foundations of Distributional Reinforcement Learning (DistRL) and Conditional Value-at-Risk (CVaR). Unlike standard value-based methods that approximate the expectation of returns, our approach learns the full distribution of returns, enabling the explicit optimization of static CVaR. We detail the theoretical framework, the limitations of naive risk-aware action selection, and the implementation of a robust trading algorithm that leverages technical indicators to make risk-adjusted decisions.

June 1, 2026
Stock Market Prediction Using FinBERT and Attention-Based Bi-LSTM
Research Paper

Stock Market Prediction Using FinBERT and Attention-Based Bi-LSTM

Predicting stock market movements is a challenging problem because prices are influenced by supply and demand, which fluctuate based on company earnings, economic conditions, investor sentiment, and global news. Traditional mathematical models often fail to capture these diverse factors. This paper presents a robust deep learning framework designed to bridge this gap by combining financial sentiment analysis with advanced time-series modeling. Our approach integrates FinBERT, a model that understands financial language, with a Bidirectional LSTM network that identifies price trends, using a special statistical technique to filter out market noise. The experimental results are promising; the strategy achieved a 62.61% Win Rate and a Sharpe Ratio of 1.80 in out-of-sample testing. These findings conclude that fusing news analysis with deep learning provides a significant advantage in forecasting short-term price movements compared to traditional methods.

June 1, 2026
Beyond Black Scholes - Local Volatility and Stochastic Volatility
Research Paper

Beyond Black Scholes - Local Volatility and Stochastic Volatility

This research report presents an exhaustive examination of advanced volatility modeling frameworks within modern quantitative finance, specifically addressing the transition from constant volatility assumptions to local volatility (LV) and stochastic volatility (SV). Motivated by the persistent empirical failure of the Black–Scholes–Merton (BSM) model to capture the implied volatility skew and smile observed in equity markets, this document provides the theoretical understanding, numerical implementation, and empirical performance of industry-standard models. We rigorously analyze Dupire’s local volatility model and Heston’s stochastic volatility model. Special emphasis is placed on accurate pricing of exotic derivatives such as barrier and vanilla options. The report details robust calibration protocols using Monte Carlo and Fast Fourier Transform (FFT) methods.

June 1, 2026
Hedging options under transaction costs
Research Paper

Hedging options under transaction costs

We study the problem of hedging European options in discrete-time markets with transaction costs, where the classical Black–Scholes replication argument fails due to the infinite cost of continuous rebalancing. Framing hedging as an optimization problem, we evaluate three strategies — full delta hedging, no-trade band hedging, and partial hedging — under a joint objective that penalizes both hedging error and cumulative transaction costs. Numerical simulations under geometric Brownian motion demonstrate that full delta hedging, while minimizing instantaneous exposure, incurs prohibitive transaction costs and yields the worst risk-adjusted performance. No-trade band strategies achieve the lowest combined objective by trading selectively only when the hedge ratio exits a prescribed tolerance region, while partial hedging offers a flexible intermediate solution by accepting residual risk in exchange for reduced trading intensity. These results establish that effective real-world hedging is fundamentally an optimization problem balancing risk reduction against cost efficiency, rather than an exercise in perfect replication.

June 1, 2026
multichannel transformer for high frequency LOB dynamic
Research Paper

multichannel transformer for high frequency LOB dynamic

This whitepaper investigates the application of Multi-Channel Transformer architectures for forecasting Limit Order Book (LOB) dynamics in high-frequency trading (HFT) environments. We propose a spatial-temporal attention mechanism that treats price and volume as distinct input channels to capture complex market microstructure. To extend prior Transformer-based LOB forecasting methods, we integrate a Liquid Neural Network (LNN) extension, replacing standard feed-forward or pooling layers with continuous-time differential equation neurons. Our implementation on the FI-2010 dataset demonstrates that liquid integration significantly improves predictive robustness at the k = 100 prediction horizon. Experimental results achieve an Accuracy of 56.94% and a Macro F1-score of 0.5681, outperforming traditional CNN-LOB baselines and demonstrating superior stability during market regime shifts. The proposed architecture demonstrates competitive predictive performance while maintaining inference latency compatible with real-time high-frequency trading constraints. Index Terms—Limit Order Book (LOB), High-Frequency Trading (HFT), Transformer, Continuous-Time Models, Liquid Neural Networks, Attention Mechanism, Market Microstructure.

June 1, 2026
Deep Weighted Monte Carlo a Hybrid Option Pricing Framework
Research Paper

Deep Weighted Monte Carlo a Hybrid Option Pricing Framework

We introduce a hybrid derivative pricing framework that combines variational deep learning with entropy-regularized Weighted Monte Carlo methods. Implied volatility surfaces are embedded into a low-dimensional latent space via a Variational Autoencoder, whose KL-regularized training objective enforces continuity and geometric regularity. A neural weight decoder then maps latent market states to probability weights over a fixed ensemble of Monte Carlo paths, yielding a discrete approximation to the market-implied risk-neutral measure. The joint calibration objective incorporates price-matching constraints, a martingale penalty, and entropy regularization, ensuring theoretical consistency with the fundamental theorem of asset pricing while preserving numerical stability. Empirical results on synthetic implied volatility surfaces demonstrate that the VAE-based framework outperforms a deterministic autoencoder baseline in both surface reconstruction fidelity and option pricing accuracy, and that the learned latent space exhibits economically meaningful market regime structure.

June 1, 2026
trial
Research Paper

trial

trial

May 29, 2026
Quant Club Logo

QUANT CLUB

Bridging the gap between financial theory and practical implementation through cutting-edge algorithms, machine learning, and quantitative analysis.

quantclubiitkgp@gmail.com

IIT Kharagpur, West Bengal

Content

  • Blog
  • Whitepapers
  • Docs

About

  • Our Team
  • Contact Us

Copyright © 2026 Quant Club IIT Kharagpur. All rights reserved.

Σ