Enhance Prediction Performance Research Articles

A hybrid stock prediction method based on periodic/non-periodic features analyses

Stock investment is an economic activity characterized by high risks and high returns. Therefore, the prediction of stock prices or fluctuations is of great importance to investors. Stock price prediction is a challenging task due to the nonlinearity and high volatility of stock time series. Existing deep learning models may not capture the periodic and non-periodic features of stock data effectively. In this paper, we propose a novel model that leverages Complete Ensemble Empirical Mode Decomposition (CEEMD), Time2Vec, and Transformer to better capture and utilize various patterns in stock data for enhanced prediction performance, and we call it ETT. CEEMD decomposes the stock data into different frequency components based on their intrinsic scales. Time2Vec provides a time vector representation that captures both periodic and non-periodic patterns while being invariant to time scaling. Transformer learns the long-term dependencies and global information from the data. We apply ETT to predict stock prices in the Chinese A-share market and compare it with several baseline models. The results show that ETT reduces the mean squared error (MSE) by an average of 4% and increases the average cumulative return by 58% on the CSI 100 and Hushen 300 datasets.

A Wavelet Decomposition Method for Estimating Soybean Seed Composition with Hyperspectral Data

Soybean seed composition, particularly protein and oil content, plays a critical role in agricultural practices, influencing crop value, nutritional quality, and marketability. Accurate and efficient methods for predicting seed composition are essential for optimizing crop management and breeding strategies. This study assesses the effectiveness of combining handheld spectroradiometers with the Mexican Hat wavelet transformation to predict soybean seed composition at both seed and canopy levels. Initial analyses using raw spectral data from these devices showed limited predictive accuracy. However, by using the Mexican Hat wavelet transformation, meaningful features were extracted from the spectral data, significantly enhancing prediction performance. Results showed improvements: for seed-level data, Partial Least Squares Regression (PLSR), a method used to reduce spectral data complexity while retaining critical information, showed R2 values increasing from 0.57 to 0.61 for protein content and from 0.58 to 0.74 for oil content post-transformation. Canopy-level data analyzed with Random Forest Regression (RFR), an ensemble method designed to capture non-linear relationships, also demonstrated substantial improvements, with R2 increasing from 0.07 to 0.44 for protein and from 0.02 to 0.39 for oil content post-transformation. These findings demonstrate that integrating handheld spectroradiometer data with wavelet transformation bridges the gap between high-end spectral imaging and practical, accessible solutions for field applications. This approach not only improves the accuracy of seed composition prediction at both seed and canopy levels but also supports more informed decision-making in crop management. This work represents a significant step towards making advanced crop assessment tools more accessible, potentially improving crop management strategies and yield optimization across various farming scales.

Freight Volume Prediction for Logistics Sorting Centers Using an Integrated GCN-BiLSTM- Transformer Model

In this paper, a comprehensive learning model is proposed for predicting cargo volume at sorting centers, which integrates GCN (Graph Convolutional Network), BiLSTM (Bidirectional Long Short-Term Memory Network), ARIMA (Autoregressive Integrated Moving Average model), and Transformer models. To achieve this, a directed weighted graph is constructed considering the transport network and average cargo volume of each sorting center. The GCN model is employed to extract spatial features from the transport connection information of the sorting centers and these features are then fed into the BiLSTM network. The BiLSTM network leverages bidirectional information flow to learn the temporal characteristics of the data. Subsequently, the GCN-BiLSTM model, which combines spatial and temporal features, is used to predict the daily cargo volume for the next 30 days. The results demonstrate that the GCN-BiLSTM model and the ARIMA-BiLSTM integrated model significantly enhance prediction performance compared to single-model approaches.

Evaluating Machine Learning Techniques for Credit Risk Management: An Algorithmic Comparison

The evaluation of credit risk has become an indispensable element within the financial sector. This research aims to conduct a comparative examination of several machine learning model's performance in predicting credit risk. This research uses comprehensive metrics to give a comparative examination of six machine learning models, including Random Forests (RF) and Support Vector Machines (SVM). The features used in the training of these models were screened by a combination of Random Forest feature importance and Recursive Feature Elimination (RFE) to ensure model accuracy. After comparing the model results, the study concluded that the Random Forest model combined with RFE performed the best among all the risk columns with an accuracy of 0.71. KNN was the next best with an accuracy of 0.69. Logistic regression was the worst performer among the six models with an accuracy of only 0.29. In the study of this paper, the imbalance of the dataset categories resulted in a weak identification of moderate risk categories. It shows that the model is not well adapted to the dataset with imbalanced categories. The paper validates the viability of machine learning in credit risk by offering useful advice on how it may be applied. To further enhance prediction performance, future studies could investigate the combination of more advanced data-balancing strategies and deep learning approaches.

Developing novel spectral indices for precise estimation of soil pH and organic carbon with hyperspectral data and machine learning.

Accurate soil pH and soil organic carbon (SOC) estimations are vital for sustainable agriculture, as pH affects nutrient availability, and SOC is crucial for soil health and fertility. Hyperspectral imaging provides a faster, non-destructive, and economical alternative to standard soil testing. The study utilizes imaging spectroscopic data from the Africa Soil Information Service (AfSIS) and Land Use and Coverage Area Frame Survey (LUCAS-2009) hyperspectral datasets, capturing spatially distributed spectral information. Machine learning (ML) approaches using high-dimensional spectral bands can be computationally expensive, while those using spectral indices are typically limited to multispectral data. This study addresses these challenges by comparing soil pH and SOC prediction using ML models, with both existing spectral indices and individual hyperspectral bands as input features. Results demonstrate that hyperspectral bands outperform existing indices in predictive accuracy, with R values ranging from 0.8 to 0.94 for both soil pH and SOC. To further enhance prediction performance, this study proposes novel spectral indices-soil pH index (SPI) and soil organic carbon index (SOCI)-specifically designed for hyperspectral data using principal component analysis (PCA) and artificial neural networks (ANN). The proposed SPI and SOCI indices address multicollinearity issues and high dimensionality in raw spectral bands, significantly improving predictive accuracy. The SPI and SOCI indices achieve R values of 0.86 for AfSIS soil pH, 0.945 for LUCAS-2009 soil pH, 0.952 for AfSIS SOC, and 0.963 for LUCAS-2009 SOC. These results show that the proposed spectral indices provide a practical solution for precision agriculture, enhancing soil pH and SOC estimations.

A Moroccan soil spectral library use framework for improving soil property prediction: Evaluating a geostatistical approach

A soil spectrum generated by any spectrometer requires a calibration model to estimate soil properties from it. To achieve best results, the assumption is that locally calibrated models offer more accurate predictions. However, achieving this higher accuracy comes with associated costs, complexity, and resource requirements, thus limiting widespread adoption. Furthermore, there is a lack of comprehensive frameworks for developing and utilizing soil spectral libraries (SSLs) to make predictions for specific samples. While calibration samples are necessary, there is the need to optimize SSL development through strategically determining the quantity, location, and timing of these samples based on the quality of the information in the library. This research aimed to develop a spatially optimized SSL and propose a use-framework tailored for predicting soil properties for a specific farmland context. Consequently, the Moroccan SSL (MSSL) was established utilizing a stratified spatially balanced sampling design, using six environmental covariates and FAO soil units. Subsequently, various criteria for calibration sample selection were explored, including a spatial autocorrelation of spectra principal component (PC) scores (spatial calibration sample selection), spectra similarity memory-based learner (MBL), and selection based on environmental covariate clustering. Twelve soil properties were used to evaluate these calibration sample selections to predict soil properties using the near infrared (NIR) and mid infrared (MIR) ranges. Among the methods assessed, we observed distinct precision improvements resulting from spatial sample selection and MBL compared to the use of the entire MSSL. Notably, the Lin’s Concordance Correlation Coefficient (CCC) values using the spatial calibration sample selection was improved for Olsen extractable phosphorus (OlsenP) by 41.3% and Mehlich III extractable phosphorus (P_M3) by 8.5% for the MIR spectra and for CEC by 25.6%, pH by 13.0% and total nitrogen (Tot_N) by 10.6% for the NIR spectra in reference to use of the entire MSSL. Utilizing the spatial autocorrelation of the spectra PC scores proved beneficial in identifying appropriate calibration samples for a new sample location, thereby enhancing prediction performance comparable to, or surpassing that of the use of the entire MSSL. This study signifies notable advancement in crafting targeted models tailored for specific samples within a vast and diverse SSL.

Space-time prediction of rainfall-induced shallow landslides through Artificial Neural Networks in comparison with the SLIP model

Rainfall-induced shallow landslides are expected to increase due to more intense precipitation linked to climate change. This study aims to develop an effective pixel-based tool for the space-time prediction of soil slips by combining a FeedForward Neural Network (FFN) with insights from the physically-based model SLIP (Shallow Landslide Instability Prediction). The FFN model was developed based on past events in four towns of the Emilia Apennines (Italy) from 2004 to 2014 under varying rainfall conditions. Among the key aspects analysed were the inclusion of both landslide and non-landslide days, the evaluation of two different cumulative rainfall periods (10 and 30 days), and various technical elements related to machine learning, including training approach, network topology, and activation function. A 2:1 imbalance in non-landslide/landslide pixels was implemented to enhance prediction performance. Prediction accuracy was measured using the Quality Combined Index (QCI), which combines AUROC, AUPRC, and F1-score. The best FFN model achieved a QCI of 0.85, accurately predicting non-landslides and minimizing false alarms. A comparison with SLIP showed that SLIP better captured the progressive destabilization in areas nearing instability, while the FFN provided a clearer distinction between stable and unstable zones. A successful blind prediction was demonstrated for a landslide in Compiano (November 2019), validating the model's applicability. SLIP also contributed to understanding the initial soil saturation and rainfall conditions, highlighting its potential to enhance FFN predictions in different meteorological scenarios. Although the developed pixel-based model could be utilized as is, further research is needed to enhance its application for early warning purposes in varying meteorological conditions.

Integrating deep learning model and virtual reality technology for motion prediction in emergencies

Predicting the evacuation behavior of pedestrians in emergencies is essential for ensuring public safety. Existing deep learning-based prediction models generally focus on crowd trajectories extrapolation in conventional scenarios but ignore the effect of emergencies on human behavior. Their performance has not been rigorously validated during emergency events such as fires and floodwaters. In this paper, we implement a combined solution involving a transformer-based network and virtual reality (VR) modeling. The proposed virtual reality-trained neural network incorporates diverse cues from human poses, moving paths, scenes, and emergency events to predict future trajectories. The virtual reality modeling creates diverse evacuation scenarios to enhance prediction performance. Moreover, based on the pretraining of our constructed VR dataset, the designed model can be applied to real-world human behavior prediction. The experimental results demonstrate our model’s superior accuracy in various scenarios, particularly for emergency evacuations, showcasing its ability to capture the dynamics of human behavior in safety-critical environments.

Health State Prediction of Lithium-Ion Battery Based on Improved Sparrow Search Algorithm and Support Vector Regression

The state of health (SOH) prediction of lithium-ion batteries is a pivotal function within the battery management system (BMS), and achieving accurate SOH predictions is crucial for ensuring system safety and prolonging battery lifespan. To enhance prediction performance, this paper introduces an SOH prediction model based on an improved sparrow algorithm and support vector regression (ISSA-SVR). The model uses nonlinear weight reduction (NWDM) to enhance the search capability of the Sparrow algorithm and combines a mixed mutation strategy to reduce the risk of local optimal traps. Then, by analyzing the characteristics of different voltage ranges, the charging time from 3.8 V to 4.1 V, the discharge time of the battery, and the number of cycles are selected as the feature set of the model. The model’s prediction capabilities are validated across various working environments and training sample sizes, and its performance is benchmarked against other algorithms. Experimental findings indicate that the proposed model not only confines SOH prediction errors to within 1.5% but also demonstrates robust adaptability and widespread applicability.

Advanced Short-Term Load Forecasting with XGBoost-RF Feature Selection and CNN-GRU

Accurate and efficient short-term load forecasting (STLF) is essential for optimizing power system operations. This study proposes a novel hybrid forecasting model that integrates XGBoost-RF feature selection with a CNN-GRU neural network to enhance prediction performance while reducing model complexity. The XGBoost-RF approach is first applied to select the most predictive features from historical load data, weather conditions, and time-based variables. A convolutional neural network (CNN) is then employed to extract spatial features, while a gated recurrent unit (GRU) captures temporal dependencies for load forecasting. By leveraging a dual-channel structure that combines long- and short-term historical load trends, the proposed model significantly mitigates cumulative errors from recursive predictions. Experimental results demonstrate that the model achieves superior performance with an average root mean square error (RMSE) of 53.29 and mean absolute percentage error (MAPE) of 3.56% on the test set. Compared to traditional models, the prediction accuracy improves by 28.140% to 110.146%. Additionally, the model exhibits strong robustness across different climatic conditions. This research validates the efficacy of integrating XGBoost-RF feature selection with CNN-GRU for STLF, offering reliable decision support for power system management.

A Spatial-Temporal Aggregated Graph Neural Network for Docked Bike-sharing Demand Forecasting

Predicting the number of rented and returned bikes at each station is crucial for operators to proactively manage shared bike relocation. Although existing research has proposed spatial-temporal prediction models that significantly advance traffic prediction, these models often neglect the unique characteristics of shared bike systems (BSS). Spatially, the entire bike-sharing system (BSS) experiences peak activity during morning and evening rush hours, whereas, during other periods, activity is localized to local stations, with some recording no rides, highlighting the need to distinguish between global and local spatial information across different times. Temporally, the historical riding records for each station exhibit non-stationary patterns, necessitating the analysis of both global trends and local fluctuations. Existing Graph Neural Network (GNN) approaches to predicting shared bike demand primarily capture static spatial-temporal data and fail to account for the dynamic nature of bike flows. Moreover, these studies focus on global spatial-temporal information without considering local nuances, making it challenging to capture spatiotemporal dynamics in fluctuating BSS. To address these challenges, we introduce the Spatial-Temporal Aggregated Graph Neural Network (STAGNN). Our model first constructs a dynamic adjacent matrix to describe the evolving connections between stations, followed by local and global information layers to capture spatial-temporal information from large-scale shared bike networks accurately. Our methodology has been validated through experiments on four real-world datasets, comparing it against benchmark models to demonstrate superior prediction accuracy. Additionally, we conduct extended experiments on four datasets during the morning and evening rush hours, and the results also affirm the efficacy of the STAGNN in enhancing prediction performance.

Soft sensor model for endpoint carbon content and temperature in BOF steelmaking based on just-in-time learning with multilayer structure preservation and sparse information enhancement

Ensuring precise prediction of the endpoint carbon content and temperature is paramount in controlling the endpoint of the basic oxygen furnace (BOF) steelmaking process. However, due to the frequent fluctuations in real-time blowing conditions, the conventional just-in-time learning soft sensor approach encounters challenges in effectively anticipating the blowing endpoint. To address these aforementioned issues, this research introduces a novel approach known as the pattern-oriented multilayer structure preservation and sparse information enhancement model (POMLSP-SIE). This approach involves dynamically generating a dataset with the same distribution based on the characteristics of the query sample pattern. It is combined with a multi-pathway dimension reduction model to preserve multi-view spatial geometry information while reducing data dimension. The low-dimensional embedding serves as dictionaries containing various hierarchical structural characteristics. Significant information within these dictionaries is sparsely represented to amplify its influence during the regression prediction process while diminishing the impact of less relevant information. This refinement aims to rectify the problem of static regression models being weak to adapt to changing working conditions, ultimately enhancing prediction performance. The effectiveness of the proposed method is substantiated through an experimental study utilising real converter steelmaking process data, thereby confirming its practical applicability.

Forecasting carbon futures returns using feature selection and Markov chain with sample distribution

The accurate forecasting of carbon returns is paramount for enabling informed investment decisions, promoting emissions reduction, and effectively shaping policies to combat climate change. In this paper, we propose a novel method to improve carbon returns predictability in a data-rich environment. The innovations of the model are manifested in two key dimensions: (i) a feature selection strategy based on the minimum prediction error is introduced; (ii) a novel Markov chain with sample distribution considering both prediction and parameter estimation is proposed to quantify the error information and perfect the prediction performance by error modification. Our empirical findings demonstrate that the proposed model outperforms a comprehensive array of competing models, both in point and interval forecasting of carbon returns. The results are consistently confirmed in various robustness checks. Finally, we show that the enhanced prediction performance of the proposed model is economically significant, which can help investors make favorable decisions.

Prediction of optical chaos using a multi-stage extreme learning machine with data uncertainty.

In this paper, we study the problem of predicting optical chaos for semiconductor lasers, where data uncertainty can severely degrade the performance of chaos prediction. We hereby propose a multi-stage extreme learning machine (MSELM) based approach for the continuous prediction of optical chaos, which handles data uncertainty effectively. Rather than relying on pilot signals for conventional reservoir learning, the proposed approach enables the use of predicted optical intensity as virtual training samples for the MSELM model learning, which leads to enhanced prediction performance and low overhead. To address the data uncertainty in virtual training, total least square (TLS) is employed for the update of the proposed MSELM's parameters with simple updating rule and low complexity. Simulation results demonstrate that the proposed MSELM can execute the continuous optical chaos predictions effectively. The chaotic time series can be continuously predicted for a time period in excess of 4 ns with a normalized mean squared error (NMSE) lower than 0.012. It also demands much fewer training samples than state-of-the-art learning-based methods. In addition, the simulation results show that with the help of TLS, the length of prediction is improved significantly as the uncertainty is handled properly. Finally, we verify the prediction ability of the multi-stage ELM under various laser parameters, and make the median boxplot of the predicted results, which shows that the proposed MSELM continues to produce accurate and continuous predictions on time-varying optical chaos.

A multi-task spatio-temporal fully convolutional model incorporating interaction patterns for traffic flow prediction

Previous traffic flow prediction studies have utilized spatio-temporal neural networks combined with the multi-task learning framework to seek complementary information for enhancing prediction performance. However, the existing methods still face two challenges: they fail to capture global interaction patterns between regions and lack consideration for inter-correlations within interaction patterns. To solve these issues, we propose a novel multi-task spatio-temporal fully convolutional model named MSTFCM. First, the model includes the interaction tensor and raster tensor as task inputs, where the interaction tensor extends the raster tensor by incorporating global interaction patterns between regions. Second, a multi-task framework combined spatio-temporal convolutional block was used to learn generalized features and interaction features. A channel spatio-temporal attention is added to adaptively adjust feature weights and capture inter-correlations. To train the MSTFCM, the uncertainty loss was designed as the learnable loss functions, which capture various flow fluctuations, to facilitate multi-task optimization. The proposed model was validated on two real-world traffic datasets collected in Xiamen, China. Experimental results showed that MSTFCM outperformed nine baselines in one-step and multi-step prediction, with slower performance degradation as predicted time intervals and steps increased. We further validated the model’s effectiveness through designed variants and visualization results.

Prediction and optimization of surface waviness of WAAM components using a hybrid Rank-Gaussian PSO algorithm and ANN

Surface waviness is crucial for the quality of components produced via the wire and arc additive manufacturing (WAAM) process. This study presents a novel method for predicting surface waviness, employing an advanced model to optimize process parameter configurations using a combination of Rank-Gaussian particle swarm optimization (RGPSO) and an Artificial Neural Network (ANN). The novelty of this process is that the RGPSO not only optimizes the hyperparameters of the ANN model to enhance prediction performance, but also addresses surface waviness optimization. The RGPSO algorithm's optimization performance is evaluated using 23 benchmark functions, demonstrating competitiveness against nine comparative meta-heuristic algorithms. Experimental data on surface waviness from the literature are utilized to train, test and validate three different prediction models, including a standalone ANN model, a PSO optimized ANN (PSO-ANN) model, and an RGPSO optimized ANN (RGPSO-ANN) model. The results indicate that the developed RGPSO-ANN model achieves the highest accuracy in terms of the metrics RMSE (0.019), R (0.996), R20.990,MAE (0.013), RMSLE0.013,and MAPE (3.46 %). It performs better than the PSO-ANN model (RMSE 0.026, R 0.975, R20.982,MAE 0.019, RMSLE0.019,and MAPE 5.73 %), and better than the ANN (RMSE 0.046, R0.991, R20.944,MAE 0.034, RMSLE0.032 and MAPE 9.31 %). The RGPSO, PSO and other optimization algorithms are then applied to minimize the surface waviness of a WAAM component. RGPSO achieved the optimal value (0.1631 mm), which corresponds to a 12.6 % reduction compared to the best value obtained using PSO.

Short-term multi-energy consumption forecasting for integrated energy system based on interactive multi-scale convolutional module

Accurate consumption forecasting is of great importance to grasp the energy consumption habits of consumers and promote the stable and efficient operation of integrated energy system (IES). To this end, this paper proposes an interactive multi-scale convolutional module-based short-term multi-energy consumption forecasting method for IES. Firstly, based on multi-scale feature fusion and multi-energy interactive learning, a novel interactive multi-scale convolutional module is proposed to extract and share the coupling information between energy consumption from different scales without increasing network parameters. Then, a short-term multi-energy consumption forecasting method is proposed, where different forecasting network structures are selected in different seasons to make full use of seasonal and coupling characteristics of the energy consumption, thus enhancing prediction performance. Furthermore, a Laplace distribution-based loss function weight optimization method is proposed to dynamically balance the loss magnitude and training speed of joint forecast tasks more robustly. Finally, the effectiveness and superiority of the proposed method are verified by comparative simulation experiments.

Machine learning-based prediction of indoor thermal comfort in traditional Chinese dwellings: A case study of Hankou Lifen

Traditional residential dwellings as architectural heritage holds substantial research value and significance for regional studies. This study investigates the indoor thermal comfort of traditional “Hankou Lifen” dwellings compared to new residences during summer through field measurements and questionnaires. The results demonstrate that the indoor thermal environment of “Lifen” houses is inferior, leading to lower thermal comfort for the residents. A linear regression model assesses summer residents' thermal comfort, considering psychological perspectives to calculate natural and acceptable temperature ranges. However, these models primarily rely solely on temperature and overlook other environmental and personal factors, due to their lack of accuracy in capturing complex non-linear relationships. To address these limitations, six machine learning models with high predictive performance are selected for analyses. Data preprocessing techniques, such as upsampling, and hyper-parameter optimization are employed to enhance prediction performance. The selection of algorithms is based on metrics such as accuracy, precision, recall, and F1 score, with further performance evaluation conducted using ROC curves and confusion matrices. Through evaluation, XGBoost and LightGBM show superior performance in predicting residential thermal comfort.Finally, SHAP(SHapley Additive exPlanations) values and PDP(Partial Dependence Plot) are used to interpret the models, revealing various factors influencing residents' thermal sensations in both new and old dwellings. These findings emphasize the significance of thermal comfort in the preservation and revitalization of historic dwellings and propose a novel framework for advancing the sustainable utilization of traditional dwellings.

Forecasting energy consumption with a novel ensemble deep learning framework

Energy baseline development is a central and significant component of an energy management system. In commercial buildings where a wide variety of internal and external variables might impact the electricity usage, having a robust forecasting method is elemental to evaluate and enhance the energy performance over time. Due to extensive progress in data analytics, a number of single prediction methods have been developed recently. Nevertheless, either development of the theoretical study or application of ensemble learning methods to enhance prediction performance and stability find lacking in energy management of commercial sectors. This research proposes a novel heterogeneous ensemble model framework for short-term electricity consumption prediction under the hourly and daily granularity to bridge this research gap. The framework is composed of two levels. The first uses five stand-alone deep learning models including Recurrent Neural Network (RNN), Long Short-term Memory (LSTM), Bidirectional-LSTM (Bi-LSTM), Gated Recurrent Unit (GRU) and Convolutional Neural Network (CNN) as base models. In contrast, the second level applies the predicted outputs of the base models as feature inputs of k nearest neighbors (KNN), which is a strong model. The Grid Search is also used to search for the key optimal hyper parameters of the base models and the strong model. To demonstrate the developed model's effectiveness, four different space types of real-world educational buildings are used for verification. The extensive database has been setup for this study to collect data from Jan. 2018 to Dec. 2022. The quantitative analysis of the model was conducted in contrast with the base models and prevalent machine learning methods, including artificial neural network (ANN), KNN, support vector regression (SVR), AdaBoost, and linear regression (LR) during the testing phase. The prediction accuracy has improved substantially, with roughly a 95 % reduction in CV-RMSE compared to that of AdaBoost and SVR. Moreover, the results signify the introduced model outperforms the base models and the ML methods in all studied buildings, thereby providing a foundation for developing effective management policies.

Exploring the potential of routine serological markers in predicting neurological outcomes in spinal cord injury

Spinal cord injury (SCI) is a rare condition with a heterogeneous presentation, making the prediction of recovery challenging. However, serological markers have been shown to be associated with severity and long-term recovery following SCI. Therefore, our investigation aimed to assess the feasibility of translating this association into a prediction of the lower extremity motor scores (LEMS) at chronic stage (52 weeks after initial injury) in patients with SCI using routine serological markers.Serological markers, assessed within the initial seven days post-injury in the observational cohort study from the Trauma Hospital Murnau underwent diverse feature engineering approaches. These involved arithmetic measurements such as mean, median, minimum, maximum, and range, as well as considerations of the frequency of marker testing and whether values fell within the normal range. To predict LEMS scores at the chronic stage, eight different regression models (including linear, tree-based, and ensemble models) were used to quantify the predictive value of serological markers relative to a baseline model that relied on the very acute LEMS score and patient age alone.The inclusion of serological markers did not improve the performance of the prediction model. The best-performing approach including serological markers achieved a mean absolute error (MAE) of 6.59 (2.14), which was equivalent to the performance of the baseline model. As an alternative approach, we trained separate models based on the LEMS observed at the very acute stage after injury. Specifically, we considered individuals with an LEMS of 0 or an LEMS exceeding zero separately. This strategy led to a mean improvement in MAE across all cohorts and models, of 1.20 (2.13).We conclude that, in our study, routine serological markers hold limited power for prediction of LEMS. However, the implementation of model stratification by the very acute LEMS markedly enhanced prediction performance. This observation supports the inclusion of clinical knowledge in the modeling of prediction tasks for SCI recovery. Additionally, it lays the path for future research to consider stratified analyses when investigating the predictive power of potential biomarkers.