A Causal Perspective of Stock Prediction Models

The stock market is one of the most important parts of the global economy, and predicting its movements has long been a challenge. Traditional methods relied on mathematical models and historical data, but advances in computing have introduced machine learning and deep learning approaches. This paper reviews four common methods used in stock prediction: Random Forest, XGBoost, Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks. Each method has strengths, such as Random Forest’s stability, XGBoost’s strong pattern recognition, CNN’s ability to process visual patterns, and LSTM’s capability to capture temporal dependencies. However, challenges remain, including limited interpretability, poor generalization to new markets, and reliance on single-modality data. To address these issues, future research should focus on combining machine learning with expert systems for better interpretability, applying domain adaptation for cross-market generalization, and using multimodal learning to integrate numerical, textual, and sentiment data. These strategies may help improve the feasibility of machine learning in real-world stock market applications.

These days, with the fast-paced development of the economy, a growing number of individuals have started making stock market investments, aiming to earn more money with wise decisions. However, a high level of risk has become inevitable because of the high level of stock price volatility, which is impacted by several variables. As a result, forecasting stocks has emerged as a critical component of financial research. One of the fundamental stock prediction models, Long Short-Term Memory (LSTM), is introduced in this article. With the consideration of the increasing level of requirement of the stock prediction, this dissertation also discusses the different ways of optimization of the LSTM models from hybridization, genetic algorithm, attention mechanism to the variant LSTM. Although the optimized models have been experimented with to show better performance in stock prediction, there are still several limitations waiting to be improved in the future. This dissertation has shed light on the analysis and comparison of stock prediction models based on the LSTM, with paramount significance to the future development of stock prediction.

Estimation of forest biomass stocks in vast and heterogeneous mountain ranges is critical in the context of climate change mitigation and remains challenging because of limited field observations and unknown relationships between variation in forest biomass and environmental heterogeneity. We addressed this challenge by using forest inventory plot observations and a novel spatial modelling approach. In the first step of our approach, we employ a rigorous clustering process to identify a homogeneous group of locations based on tree species and topoclimatic variables and predict potential forest aboveground biomass (AGB). Subsequently, in the second step, we incorporate finer-scale variables, including proxies of forest structure, disturbance likelihood, and elevation zones, to model deviations from the predicted potential AGB. Our method significantly improves forest AGB estimation in heterogeneous mountain landscapes, achieving a 25% reduction in prediction error compared to the best-performing existing model. The final forest AGB map, generated at 30m resolution, reveals distinct spatial patterns, with the Central Himalayas emerging as a key carbon reservoir, harbouring forest patches exceeding 1000 t ha-1. Aggregation of these predictions yielded a total forest AGB of 1982 Mt. In addition, we produced a 250m resolution potential forest AGB map with associated prediction standard error. The spatially explicit estimates of actual and potential forest biomass presented is important step towards elucidation of spatial distribution patterns of forest carbon pools and environmental controls. It also provides support for critical initiatives, including climate change mitigation strategies, monitoring forest landscape restoration, and combatting forest degradation challenges. The proposed approach, integrating both broad-scale environmental controls and fine-scale deviations, offers a robust method that is potentially applicable other mountainous regions and contributes for tracking changes in forest carbon over time, essential for REDD+ initiatives.

This research aims to make a comprehensive comparative analysis between traditional machine learning methods and Long Short-Term Memory networks (LSTMs) towards stock return predictions. By using the full Kaggle market data set (the Stock Price Prediction Challenge), we managed to generate an integrated forecasting pipeline for stock prediction. In this data set, we used 45 stocks and three major indexes to engineer extensive features and a strong model validation. After considering all factors, Gradient Boosting, outperforming both traditional methods and LSTM, achieves the greatest training performance: Mean Squared Error (MSE) of 0.000135, R2 of 0.027195, and mean absolute percentage error (MAPE) of 155.85%. Contrary to the initial assumption, all models exhibited severe overfitting. A significant performance drop on the validation set suggests a major challenge in practical prediction use. The findings indicate that while the models do not provide practically useful, accurate return predictions based solely on price information, they do provide strong comparison benchmarks and methodological suggestions in future studies involving other data and stronger regularisation approaches.

Abstract. In Pacific Rim regions highly exposed to climate variability, accurate projections of soil organic carbon (SOC) are critical for future effective land management and climate adaptation strategies. This study integrated digital soil mapping with CMIP6-based climate correlative spatial modelling to estimate the spatiotemporal distribution of SOC stocks in subtropical (Zhuoshui River) and tropical (Laonong River) watersheds in Taiwan. We collected 901 soil samples and data on 18 environmental covariates and modeled SOC stocks at a 20 m resolution through the Cubist and random forest algorithms, which were also combined with regression kriging. The Cubist-based kriging model was discovered to achieve the highest performance in SOC stock prediction. Mountainous areas were found to contain >80 % of SOC stocks, and tropical zones were discovered to store substantially less carbon than subtropical zones. The space-for-time estimates derived from future climate analogues indicated considerable spatial heterogeneity in potential steady-state SOC conditions. Under SSP1-2.6, climatic analogues associated with cooler and drier conditions corresponded to lower SOC stocks – up to 20.9 % lower than baseline – particularly in uplands, whereas SSP2-4.5 analogues were associated with SOC states that were 7.9 % higher, especially in mountainous regions. These contrasts reflect spatial associations observed in the contemporary landscape rather than mechanistic predictions of erosion, productivity, or carbon-cycle responses. Partial least squares path modeling revealed a strong climate–topography interaction and explicitly quantified their contributions to SOC stocks, dominated by topography and followed by prolonged dry spells (CDD). This interaction is more pronounced in uplands than in mountainous areas, where topography mitigates temperature extremes and their effects on SOC retention. Extended CDD may decrease organic inputs by reducing vegetation growth and soil moisture, thereby enhancing carbon losses. Examining the interactions between climatic extremes, landscape types, and SOC stocks is essential for enhancing soil resilience and ensuring stable SOC stocks in the future.

The mechanism and prediction of stock market volatility constitute a long-standing core research focus within financial engineering. Machine learning stands out due to its exceptional capacity for modeling nonlinear structures and has consequently become a principal technique in stock price prediction. It effectively uncovers complex market interrelations and underlying trends. A common limitation in existing research, however, lies in inadequate model interpretability and a lack of integrated visualization support. To bridge these gaps, this research selects Vanke A (000002)—a representative real estate firm—as the case study, builds a stock prediction model using the random forest algorithm, embeds visualization features, and ultimately develops an end-to-end decision support system. Empirical findings reveal that the model successfully tracks short-term stock price fluctuations and exhibits strong predictive consistency and robustness in stable market climates. Further analysis of feature importance underscores the persistent influence of volume-price indicators—such as opening price, historical trading volume, and turnover rate—as pivotal drivers of stock price movement, reaffirming the dominant role of transactional data in short-term forecasting. Theoretically, this work not only corroborates the efficacy of ensemble learning in modeling financial time series but also accentuates the critical role of visualization tools in enhancing model transparency and supporting investment decisions.

Artificial intelligence (AI) has become crucial in quantitative finance, driving improvements in investment modeling, strategy development, and risk assessment. This paper reviews recent advancements in AI-enhanced quantitative investing, emphasizing deep learning, reinforcement learning, and algorithmic optimization. Research by Ding and Qin achieved over 97% accuracy in stock predictions using Long Short-Term Memory (LSTM) networks, while Zhang's XGBoost framework improved high-frequency trading performance, yielding a mean squared error (MSE) of 0.1918. Xu et al. implemented a Kalman Filter-based PI clock servo, achieving nanosecond-level synchronization precision (59.37 ns) in multi-hop networks, enhancing scalability. In financial risk forecasting, Duan et al. increased K-Means clustering accuracy to 99.4%, reducing false alarm rates by nearly 48%. Additionally, optimization techniques like PPO, CVaR-based risk management, and quantum optimization have improved stability, risk control, and computational efficiency. However, challenges like data diversity, model interpretability, and regulatory oversight persist. In conclusion, integrating AI into quantitative finance enhances predictive accuracy and decision-making, transforming intelligent asset management amid changing risk landscapes.

Spatial prediction of soil organic carbon stocks across contrasting Andean basins, Peru

Complex system and PS-LSTM prediction of cryptocurrencies, stocks, bonds, exchange rates and commodities

Bayesian spatial prediction of soil organic carbon stocks in eastern DRC using INLA-SPDE and environmental covariates

Optimizing mean conditional value-at-risk portfolios through deep neural network stock prediction

LSTM stock prediction model based on blockchain

Long Short-Term Memory (LSTM) has long been a popular prediction model across multiple fields, and extensive research has been conducted on its applications. Among these studies, most employ the traditional traintest split ratio for experiments. However, in the field of stock prediction, this ratio-based split is not practical. From an investment perspective, a key consideration lies in determining the optimal length of historical data required to produce robust and accurate predictions for future market periods. To address the practical utilization of prediction models, this study examines how external parameterssuch as time window, training size, and test size (in terms of time length)affect the performance of the LSTM model in predicting the stock price of The Coca-Cola Company. The dataset includes the historical Open, Low, High, and Close prices from January 1, 2000, to December 31, 2024, with the Close price serving as the target variable. Model performances under different parameter settings are compared using Mean Squared Error (MSE) and prediction plots. The experimental results provide insights into how these external parameters influence the LSTMs performance in stock prediction.

Within the broader scope of applications of artificial intelligence to financial forecasting, this paper studies how to deal with nonlinearity and high noise for stock price prediction. We apply BPNN to investigate appropriate configuration settings for long-term stock prediction. Through systematically adjusting hyperparameters like the data span, the hidden-layer structure, or the number of neurons, we show that the most appropriate model for capturing Nestl stocks long-run trend characteristics is the one trained on 10 years of data, with 16 neurons over two hidden layers. This setting is not only an inherent regularity of a mature companys behavior, but it also has robustness as we have done many comparisons. It makes it part of using neural networks in an attempt to achieve financial forecasts.

This paper investigates the predictability of stock returns in the Chinese market through the lens of consumption–wealth dynamics within a broader financial system. We focus on two key state variables derived from modern consumption-based asset pricing models: the ratio of log surplus consumption (scr), from the habit-formation framework, and the log consumption–wealth ratio (cay), from the long-run cointegration framework. Using quarterly data from the CSI 300 index between 2012Q1 and 2018Q4, our system-based analysis reveals a horizon-dependent pattern of predictability. The results show that scr is a strong short-term predictor of excess stock returns, reflecting cyclical changes in risk aversion, whereas cay demonstrates superior predictive power over mid- to long-term horizons, consistent with its role as a proxy for long-run expectations. Interestingly, combining scr and cay does not improve predictive performance, suggesting that the economic mechanisms they capture are distinct rather than complementary in the Chinese market. These findings provide evidence on how interconnected macro-financial variables shape stock return dynamics, highlighting the importance of considering temporal horizons when modeling financial systems.

Gradient-Aligned Dual Encoder Framework for Stock Prediction Combining Self-Supervised Contrastive Learning and Regression

Background: Accurate prediction of real estate stock prices plays a significant role in enabling investment choices, econometric forecasting, and strategic investment decisions in emerging markets. The Kurdistan Region of Iraq is among the emerging markets that are marked by non-homogeneity of data, limited representativeness, and non-availability of standardized appraisal methods. Aims: This study aims to develop and test a stable hybrid machine learning model for predicting real estate stock prices in the Kurdistan region of Iraq. The new approach highlights methodological quality, representativeness of the dataset, and replicability in a bid to fill the gaps observed in the existing literature. Methodology: The research makes use of the Kurdistan House Price Prediction (KHPP) data, which has various property features. Preprocessing of the data involved leakage-secure feature encoding, missingness checks, and deduplication verification. A combined model of Random Forest (RF) and Extreme Gradient Boosting (XGB) was used, alongside nested cross-validation for hyperparameter selection and testing. Statistical significance testing, bootstrapped confidence intervals, and feature ablation experiments were also performed for redundancy. Results: The findings further indicate that the hybrid RF+XGB model outclassed individual models according to Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). While performance differences remained minute, the hybrid model always returned stable results per fold. Evaluation also observed challenges interpreting R2 and emphasized careful reporting for the raw target scale. The study also demonstrated an externally available benchmarking plan using the Ames dataset for cross-national comparability. Conclusion: The new hybrid approach provides a systematic, replicable, and efficient framework for real estate stock price prediction within the Kurdistan Region of Iraq. By adding robust validation protocols and statistical testing, this study enhances belief in hybrid ensemble methods for under-represented markets. Future research must move beyond single-modal data and local criteria for greater generalizability and real-world applications.

ABSTRACT Stock prediction is essential for informed investment decisions, risk management, and economic planning. This study aims to forecast stock price movements across the top eight Indian sectors using a combination of four multidimensional input types – numerical data, technical patterns, technical indicators, and textual data – within a unified predictive framework. The analysis focuses on eight major Indian sectors: Auto, Bank, Financial Services, FMCG, Metal, IT, PSU Bank, and Realty, over the period from 2017 to 2021. The study employs deep learning models including LSTM, GRU, CNN, Bi-LSTM, and a hybrid LSTM-GRU architecture on the integrated dataset. Model performance is evaluated using RMSE, MdSE, MAE, and R2. Among the models tested, the hybrid LSTM-GRU model achieved the highest prediction accuracy, with a peak performance of 98% in the IT sector. Studies that combine all four input types within a single predictive framework are rare, and the integration of an extensive range of technical patterns along with sector-wise analysis is seldom addressed in the existing literature. This study offers a more comprehensive and multidimensional approach to stock prediction, contributing valuable insights for investors, researchers, and policymakers.

Wider implementation of circular economy practices in the construction sector requires predicting the availability, properties, and qualities of reusable components beyond mere material quantities. However, data describing the geometric attributes, material properties, and modes of assembly of products available in current buildings is limited. To address this data gap, we present the Digital Inventory for Swiss Construction Systems (DISCS), a database structure that provides detailed information on component attributes in as-built load-bearing and insulating layers of existing buildings. The dataset currently provides granular data for 102 buildings in Switzerland, each digitalised into a building information model and parameterised using a custom library of 78 attributes. The database structure facilitates the operationalisation of an inventory for construction systems, providing a basis for stock prediction that supports the upscaling of component reuse in new projects. This data descriptor motivates the need for such a database, describes its ontology, and validates its use through a series of first-level analyses.