Forecast Horizon Research Articles

A U-MIDAS modeling framework for forecasting carbon dioxide emissions based on LSTM network and LASSO regression

A U-MIDAS modeling framework for forecasting carbon dioxide emissions based on LSTM network and LASSO regression

Machine Learning Techniques for Cafeteria Demand Forecasting: An Institutional Case

This research investigates the application of machine learning techniques for cafeteria demand forecasting within institutional settings, addressing critical operational challenges in food service management. Using a comprehensive methodological framework, the study analyzes turnstile entry data from an academic institution across November-December 2023 to develop and evaluate three complementary forecasting models: XGBoost with time-based features, Long Short-Term Memory (LSTM) networks, and Prophet models with domain-specific components. The comparative analysis reveals differentiated performance characteristics across various forecasting dimensions, with XGBoost demonstrating superior accuracy for daily forecasting (MAE=16.23, MAPE=8.32%), LSTM excelling at high-resolution 15-minute interval prediction (MAE=5.37, MAPE=11.64%), and Prophet exhibiting greater stability for extended forecast horizons. A weighted ensemble methodology integrating these complementary approaches yields consistent performance improvements across multiple evaluation metrics, achieving 4.7% reduction in daily MAE and 3.4% reduction in 15-minute interval MAE compared to the best individual models. Feature importance analysis reveals the significance of recent historical patterns, weekly cyclical components, and academic calendar effects, validating the theoretical multi-level temporal structure of institutional demand. The operational impact assessment demonstrates substantial potential benefits, including estimated food waste reduction of 6.2%, enhanced service level maintenance, and improved resource utilization. This research contributes methodological advancements through its multi-resolution forecasting framework, systematic feature engineering approach, and context-sensitive ensemble integration methodology, while providing practical implementation guidance for institutional food service operations seeking to enhance operational efficiency and sustainability through improved demand prediction.

Volatility connectedness of commodity futures and its application in portfolio optimization

This paper investigates volatility connectedness and its application in portfolio optimization, focusing on China’s commodity futures market. Using the Time-Varying Parameter Vector Autoregression model, we examine the dynamic connectedness of commodities. In addition to realized volatility, continuous volatility is introduced to reduce distortions caused by intraday jumps. We develop new strategies by (1) incorporating connectedness penalties into the objective function and (2) imposing additional constraints on commodity weights based on net connectedness and degree centrality. These strategies are compared to traditional Markowitz and other benchmark portfolios. Our findings demonstrate that the proposed strategies significantly outperform the benchmark portfolios, particularly during crisis periods. These results are robust to alternative volatility measures, different lag lengths, extended forecast horizons, and various adjustments to key model parameters. Furthermore, the continuous volatility-based strategies outperform their realized volatility counterparts in most cases, offering enhanced stability and reduced short positions. These findings offer valuable insights into the role of connectedness in mitigating systemic risks and optimizing portfolio strategies, emphasizing the necessity of incorporating both covariance and volatility connectedness in the decision-making process.

Forecasting Market Fear: the roles of policy uncertainty and geopolitical Risk

ABSTRACT This paper examines the predictive power of Economic Policy Uncertainty (EPU) and Geopolitical Risk (GPR) indices on market fear, as reflected by implied volatility indices across various assets and regions. A GARCH-MIDAS model is employed to analyse how low-frequency economic policy and geopolitical risks affect high-frequency market fear indices, using realized volatility as a benchmark. The study incorporates global, U.S., and Russia-based EPU, alongside multiple GPR variants, to assess their influence on implied volatility across stocks, commodities, and IT equities. Our results show that GARCH-MIDAS models incorporating external uncertainty indices significantly outperform the conventional GARCH-MIDAS model based on realized volatility alone, with particularly strong performance observed for global and U.S. uncertainty measures across multiple forecast horizons. These results highlight the importance of monitoring external uncertainties to support pre-emptive policy measures and to guide investors in integrating such insights into risk assessment models for improved volatility management.

Salmon futures prices as forecasts

Futures markets serve two main purposes, risk transfer and price forecasting. Both are relevant in the case of Norwegian farmed Atlantic salmon, as the spot price is highly volatile and hard to predict. However, the salmon futures market suffers from low liquidity, thus limiting the effectiveness of risk transfer. What about price forecasting? We consider futures prices as point forecasts of the future spot price. We evaluate them by statistical optimality criteria and find a downward bias that increases with the forecast horizon at a rate of over 10% a year. Our analysis reveals no additional evidence of forecast suboptimality. The results should be of interest to decision makers who rely on salmon futures prices as point forecasts of the future spot price.

Interdecadal cycles in Australian annual rainfall

Abstract. Extremes of Australian rainfall have profound economic, ecological and societal impacts; however, the current forecast horizon is limited to a few months. This study investigates interdecadal periodicity in annual rainfall records across eastern Australia. Wavelet analysis was conducted on rainfall data from 347 sites covering 130 years (1890–2020). Prominent cycles were extracted from each site and clustered using a Gaussian mixture model. This revealed three principal cycles centred around 12.9, 20.4 and 29.1 years that were highly significant over red noise using a t test (p<0.0001). Overall, the three cycles combined had a mean contribution to the total rainfall variance (R2) of 13 % across all of the sites, but this was up to 29 % at individual sites. Both the 12.9- and 20.4-year cycles were detected at over 95 % of the sites. The strength of each cycle varied over time, and this amplitude modulation of the signal showed a systematic movement across the area investigated. Eighty-six percent of extremely wet years fell within the positive phase of the combined reconstruction, with 80 % of extremely dry years falling in the negative phase. These results indicate underlying periodicity in annual rainfall across eastern Australia, with the potential to build this into long-term forecasts. This concept has been suggested in the past but has not been rigorously tested. These findings open new paths for research into rainfall patterns in Australia and internationally. They also have broad implications for the management of water resources across all sectors.

Marine heatwaves in the Mediterranean Sea: a convolutional neural network study for extreme event prediction

Abstract. In recent decades, the Mediterranean Sea has experienced a notable rise in the occurrence and intensity of extreme warm temperature events, referred to as marine heatwaves (MHWs). Hence, the ability to forecast Mediterranean MHWs in the short term is an area of ongoing research. Here, we introduce a novel machine learning (ML) approach specifically tailored for short-term predictions of MHWs in the basin using an attention U-Net convolutional neural network. Trained on daily sea surface temperature anomalies (SSTAs) and gridded fields of MHW presence and absence between 1982–2017, our model generates a spatiotemporal forecast of MHW occurrence up to 7 d in advance. To ensure robust performance, we explore various configurations, including different forecast horizons and U-Net architectures, number of input days, features, and different subset splits of train–test datasets. Comparative analysis against a persistence benchmark reveals an improvement of 15 % in forecasting accuracy of MHW presence for a 7 d forecast horizon. We also demonstrate an improvement of MHW prediction accuracy as the forecast horizon decreases, albeit with a smaller discrepancy between the persistence benchmark, which also results in high accuracy for the 3 d forecasts. Our proposed ML methodology offers a data-driven prediction of MHWs with reduced computational requirements, which can be applied across different regions of the global ocean, providing relevant stakeholders and management authorities with essential lead time for implementing effective mitigation strategies.

Inflation Forecasting in Time Series Models Using High Frequency Data

The article examines ways to improve inflation forecasting by using high frequency consumer price data in time series models. The purpose of increasing the number of observations available at a higher frequency is to increase the accuracy of inflation forecasts. The theoretical part of the paper considers the advantages and disadvantages of using high frequency price data in ADL, VAR and MIDAS inflation models with both single and mixed data frequency. The empirical section traces out the effects of including an online price index available at a daily or weekly frequency during the period from 2020 to 2023 in the forecast model for the consumer price index. The article compares the forecasts of consumer prices by applying the VAR, MFVAR and MIDAS models which include data from a high frequency regressor with the forecasts obtained through auto-ARIMA models. The conclusion about the difference in the quality of the short-term forecast of consumer price dynamics in these models is based on the difference of the forecast error indicator of the models. The results provide some evidence that short-term out-of-sample CPI dynamic forecasting becomes more accurate when online price data is included (namely in the class of multidimensional time series models when data is included in the model at a higher frequency). However, the advantage derived from including high frequency online price data in models decreases as the forecast horizon is extended. The results show the importance of including online price data in inflation models in a disaggregated form while forecasting price trends of the nearest future.

Hybrid CNN–GRU model for hourly flood forecasting index: case studies from the Fiji islands

Abstract Developing flood forecasting techniques at short timescales improve early warning systems to mitigate severe flood risk and facilitate effective emergency response strategies at vulnerable sites. In this study, we develop a hybrid deep learning algorithm, C-GRU, by integrating Convolutional Neural Networks (CNN) with Gated Recurrent Unit (GRU) model and evaluate its effectiveness in forecasting an hourly flood index ( $$SWRI_{24-hr-S}$$ ) in five flood-prone, specific study sites in Fiji. The model incorporates statistically significant lagged $$SWRI_{24-hr-S}$$ with real-time hourly rainfall measurements obtained from rainfall stations, and comparative analysis is performed against benchmark models: CNN, GRU, Long Short-Term Memory and Random Forest Regression. The proposed model’s outputs comprise the $$SWRI_{24-hr-S}$$ predicted at each specific site at a lead time of 1-h. The results demonstrate that the proposed hybrid C-GRU model outperforms all the other models in accurately forecasting $$SWRI_{24-hr-S}$$ over a 1-hourly forecast horizon. Across all of the study sites, the proposed model consistently generates the highest r (0.996–0.999) and the lowest RMSE (0.007–0.014) and MAE (0.003–0.004) in the testing phase. The proposed hybrid C-GRU model also achieves the highest Global Performance Index (GPI) values and the largest percentage of forecast errors (FE) ( $$\approx $$ 98.9–99.9%) within smaller error brackets (i.e., $$|\hbox {FE}|< 0.05$$ ) across all study sites. Using the methodologies developed, we show the practical application of the proposed framework as a decision support system for early flood warning, demonstrating its potential to enhance real-time monitoring and early warning systems with broader application to flood-prone regions.

Solar radiation prediction: A multi-model machine learning and deep learning approach

The increasing integration of renewable energies into electrical grids necessitates accurate forecasting of meteorological variables, particularly solar irradiance. This study presents a novel long-term solar irradiance forecasting approach, utilizing meteorological data from the National Renewable Energy Laboratory spanning 1988–2022. Focusing on five input variables—solar irradiance, dew point, temperature, relative humidity, and wind speed—this study evaluates the predictive performance of 13 data-driven models, comprising ten machine learning (ML) and three deep learning (DL) algorithms. Among them, gradient boosting regressor (GBR) and recurrent neural network (RNN) emerged as top performers in ML and deep learning, respectively. In order to choose the most suitable model for the long and short term, four forecast time-horizons (1, 8, 16, and 24 h) were also taken into consideration for the accurate models. A feature selection process using Pearson’s coefficient identified the most relevant inputs, while quantile regression was employed for uncertainty assessment, mean prediction interval, and prediction interval coverage probability models. This study demonstrates that RNN excels in short-term predictions, while GBR is more effective for long-term forecasts. A new hybrid approach GBR-RNN model was developed, achieving superior performance in terms of RMSE, MAE, and R2 metrics. This multi-model approach, integrating both ML and DL techniques, enhances solar irradiance forecasting by addressing input uncertainty and considering various forecast horizons. The findings contribute to the ongoing advancement of renewable energy forecasting by providing robust, accurate, and uncertainty-aware predictive models. Moreover, this approach helps identify the best-performing model, enabling more reliable and precise solar irradiance forecasts for energy management. This highlights both the improvement in forecasting methods and the importance of selecting the best model for accuracy.

Effectiveness of three machine learning models for prediction of daily streamflow and uncertainty assessment.

This study evaluates three Machine Learning (ML) models-Temporal Kolmogorov-Arnold Networks (TKAN), Long Short-Term Memory (LSTM), and Temporal Convolutional Networks (TCN)-focusing on their capabilities to improve prediction accuracy and efficiency in streamflow forecasting. We adopt a data-centric approach, utilizing large, validated datasets to train the models, and apply SHapley Additive exPlanations (SHAP) to enhance the interpretability and reliability of the ML models. The results show that TKAN outperforms LSTM but slightly lags behind TCN in streamflow forecasting. TKAN demonstrated strong alignment with observed statistical parameters, achieving a Mean Absolute Error (MAE) of 5.799 m³/s and a Nash-Sutcliffe Efficiency (NSE) of 0.958, compared to MAE and NSE values of 8.865 m³/s and 0.942 for LSTM, and 5.706 m³/s and 0.961 for TCN, respectively. Multi-step forecasting revealed TKAN's robust performance up to a three-day forecast horizon, with a slight decline in accuracy as the forecast period extended. Uncertainty analysis indicated reasonable variance levels, with a mean 3-day forecast uncertainty of 35.02% at a 95% confidence level for TKAN, compared to 39.95% for LSTM and 28.46% for TCN. For a 7-day forecast, TKAN showed a mean uncertainty of 40.97%, compared to 45.01% for LSTM and 36.22% for TCN. By enhancing model transparency and improving datasets, this study significantly advances the integration of machine learning into hydrological forecasting, offering robust methods for developing adaptive water management systems in response to changing climate conditions.

A Comparative Evaluation of Two Bias Correction Approaches for SST Forecasting: Data Assimilation Versus Deep Learning Strategies

This study introduces two distinct post-processing strategies to address systematic biases in sea surface temperature (SST) numerical forecasts, thereby enhancing SST predictive accuracy. The first strategy implements a spatiotemporal four-dimensional multi-grid analysis (4D-MGA) scheme within a three-dimensional variational (3D-Var) data assimilation framework. The second strategy establishes a hybrid deep learning architecture integrating empirical orthogonal function (EOF) analysis, empirical mode decomposition (EMD), and a backpropagation (BP) neural network (designated as EE–BP). The 4D-MGA strategy dynamically corrects systematic biases through a temporally coherent extrapolation of analysis increments, leveraging its inherent capability to characterize intrinsic temporal correlations in model error evolution. In contrast, the EE–BP strategy develops a bias correction model by learning the systematic biases of the SST numerical forecasts. Utilizing a satellite fusion SST dataset, this study conducted bias correction experiments that specifically addressed the daily SST numerical forecasts with 7-day lead times in the Kuroshio region south of Japan during 2017, systematically quantifying the respective error reduction potentials of both strategies. Quantitative verification reveals that EE–BP delivers enhanced predictive skill across all forecast horizons, achieving 18.1–22.7% root–mean–square error reduction compared to 1.2–9.1% attained by 4D-MGA. This demonstrates deep learning’s unique advantage in capturing nonlinear bias evolution patterns.

Transformers and Long Short-Term Memory Transfer Learning for GenIV Reactor Temperature Time Series Forecasting

Automated monitoring of the coolant temperature can enable autonomous operation of generation IV reactors (GenIV), thus reducing their operating and maintenance costs. Automation can be accomplished with machine learning (ML) models trained on historical sensor data. However, the performance of ML usually depends on the availability of large amount of training data, which is difficult to obtain for GenIV, as this technology is still under development. We propose the use of transfer learning (TL), which involves utilizing knowledge across different domains, to compensate for this lack of training data. TL can be used to create pre-trained ML models with data from small-scale research facilities, which can then be fine-tuned to monitor GenIV reactors. In this work, we develop pre-trained Transformer and long short-term memory (LSTM) networks by training them on temperature measurements from thermal hydraulic flow loops operating with water and Galinstan fluids at room temperature at Argonne National Laboratory. The pre-trained models are then fine-tuned and re-trained with minimal additional data to perform predictions of the time series of high temperature measurements obtained from the Engineering Test Unit (ETU) at Kairos Power. The performance of the LSTM and Transformer networks is investigated by varying the size of the lookback window and forecast horizon. The results of this study show that LSTM networks have lower prediction errors than Transformers, but LSTM errors increase more rapidly with increasing lookback window size and forecast horizon compared to the Transformer errors.

Forecasting Inflation in Mongolia Using Machine Learning

This paper explores the significance of machine learning (ML) techniques in forecasting inflation and identifying its drivers in Mongolia, a commodity-dependent developing economy. ML methods are handy in dealing with a large database and specify flexible relationships among variables. Our empirical work resulted in several novel findings. First, we present that ML methods (XGBoost, Random forest and Ridge regression) with large datasets can produce more accurate forecasts than the standard benchmarks, particularly for longer forecast horizons. The dominance of XGBoost and Random forest longer forecast horizons indicates the existence of nonlinearities in the inflation dynamics, relevant to forecasting inflation. Second, the performance of the factor-augmented autoregressive (FAAR) model depends on the approach used in identifying the optimal number of factors. Third, the best predictors of inflation change considerably over forecast horizons. Supply factors are the best performers in predicting inflation for short-horizon, while demand-side factors are the most important factors for longer forecast horizons. Fourth, the selection of variables is quite similar across the ML methods.

Multivariate linear time-series modeling and prediction of cerebral physiologic signals: review of statistical models and implications for human signal analytics.

Cerebral physiological signals embody complex neural, vascular, and metabolic processes that provide valuable insight into the brain's dynamic nature. Profound comprehension and analysis of these signals are essential for unraveling cerebral intricacies, enabling precise identification of patterns and anomalies. Therefore, the advancement of computational models in cerebral physiology is pivotal for exploring the links between measurable signals and underlying physiological states. This review provides a detailed explanation of computational models, including their mathematical formulations, and discusses their relevance to the analysis of cerebral physiology dynamics. It emphasizes the importance of linear multivariate statistical models, particularly autoregressive (AR) models and the Kalman filter, in time series modeling and prediction of cerebral processes. The review focuses on the analysis and operational principles of multivariate statistical models such as AR models and the Kalman filter. These models are examined for their ability to capture intricate relationships among cerebral parameters, offering a holistic representation of brain function. The use of multivariate statistical models enables the capturing of complex relationships among cerebral physiological signals. These models provide valuable insights into the dynamic nature of the brain by representing intricate neural, vascular, and metabolic processes. The review highlights the clinical implications of using computational models to understand cerebral physiology, while also acknowledging the inherent limitations, including the need for stationary data, challenges with high dimensionality, computational complexity, and limited forecasting horizons.

Interpretable physics‐informed graph neural networks for flood forecasting

AbstractClimate change has intensified extreme weather events, with floods causing significant socioeconomic and environmental damage. Accurate flood forecasting is crucial for disaster preparedness and risk mitigation, yet traditional hydrodynamic models, while precise, are computationally prohibitive for real‐time applications. Machine learning surrogates, such as graph neural networks (GNNs), improve efficiency but often lack physical consistency and interpretability. This paper introduces HydroGraphNet, a novel physics‐informed GNN framework that, for the first time, integrates the Kolmogorov–Arnold Network (KAN) to enhance model interpretability in unstructured mesh‐based flood forecasting. The framework embeds mass conservation laws into the loss function, ensuring physically consistent predictions. Additionally, it employs an autoregressive encoder–processor–decoder architecture that captures spatiotemporal flood dynamics while mitigating error accumulation over long forecasting horizons. Validation on flood data from the White River near Muncie, Indiana, demonstrates a 67% reduction in prediction error, near‐zero mass balance error, and a 58% improvement in the critical success index for major flood events compared to a baseline GNN model. These results highlight the potential of the proposed framework to advance real‐time flood forecasting with improved physical consistency and interpretability.

Developing a demand planning strategy for joint forecasting and employing analytical tool in an empirical case study

Product demand prediction is highly important as levels of inventory, customers forecast accuracy and overall performance are directly impacted by this demand planning. As the main problem of this research, accurate forecasting as a crucial challenge for achieving high performance is targeted. This paper has developed a demand planning framework with analytical capabilities of sales and consumer patterns, historical sales, and seasonality data to maximize the company’s ability to satisfy consumer demand. The framework has been implemented in an empirical business environment and has been analyzed from a practical viewpoint. A comprehensive picture of all factors is acquired from this viewpoint for successful implementation. The primary objective of the framework is to enhance the demand planning process in joint forecast project and how these improvements will help other business units, a crucial component of the supply chain, to achieve greater performance levels. As a result, the process of demand planning is examined from a system viewpoint for its integration with the supply chain E2E process, performance, forecasting methods, and logistical structure which serve as the foundation for assessment of the current condition and context of the new system. A deeper focus is placed on the analysis of the demand planning processes and forecasting techniques. Several root causes are found based on the customer forecast inputs like forecast horizon, forecasting methods, standard process with customers, performance measurement and communication with customers. Integration of JDA, ERP, and other data sources, combined with flexible forecasting tools, greatly refines the demand planning while enhancing overall decision making. By blending statistical models with real-time customer insights, the firms maintain an agile, efficient supply chain that can quickly adapt to market disruptions.

Integrating Multi-Source Data for Long Sequence Precipitation Forecasting

Long-sequence precipitation forecasting is critical for both meteorological science and smart city applications. The primary objective of this task is to predict future radar echo sequences, which provide high resolution and timely references for atmospheric precipitation distribution based on current observations. However, the chaotic nature of precipitation systems poses significant challenges in extending reliable forecast horizons. Most existing methods struggle with accuracy and clarity when extended to long-sequence predictions, such as three-hour forecasts. This is primarily due to the insufficiency of spatio-temporal information within a single modality over time. In this paper, we propose a cascading forecasting framework that adaptively extracts and integrates multimodal spatio-temporal information to support accurate and realistic long-sequence radar forecasting. Our framework includes a temporal adaptive predictor and a flow-based precipitation distribution adaptor. The predictor utilizes a multi-branch encoder-decoder architecture. This design allows it to extract meteorological sequences from multiple sources at varying scales, resulting in an initial global precipitation estimate. The core component is a carefully designed cross-attention module with a temporal adaptive layer to enhance multi-modality alignment. The initial estimate is then refined by the flow-based adaptor, which adjusts the prediction to match the target precipitation distribution, enhancing local details and correcting extreme precipitation patterns. We validated our method using real multi-source dataset for long-sequence forecasting, and the experimental results demonstrate that our approach outperforms existing state-of-the-art methods.

Forecasting Chinese stock market volatility: the role of rare event shocks from US stock market

ABSTRACT This paper investigates the role of rare event shocks (RES) from the US stock market in forecasting Chinese stock market volatility. In doing so, we use the jumps in VIX to capture the RES from the US stock market, and extend the realized EGARCH-MIDAS (REGARCH-MIDAS) model to accommodate the spillover information of RES. Our empirical results show that RES positively impacts Chinese stock market long-term volatility through investor sentiment and market liquidity. The proposed REGARCH-MIDAS-RES model outperforms competing models in out-of-sample forecasting, with findings robust to DM test, R o o s 2 test, different volatility states, different forecast horizons, Granger causality test, an alternative proxy for RES, and an alternative emerging market. Furthermore, incorporating RES into a volatility-timing strategy generates significant economic gains for a mean-variance utility investor. This finding is robust to different risk-aversion levels.

A transformer-based method for correcting daily SST numerical forecasting products

This study introduces applies a Transformer-based method to correct daily Sea Surface Temperature (SST) numerical forecasting products, addressing persistent challenges in short-term SST prediction. The proposed approach utilizes a Transformer model architecture to capture complex spatiotemporal dependencies in SST error fields, enabling efficient prediction of forecast errors across multiple time scales. The method was applied to SST hindcast data from the First Institute of Oceanography (FIO-COM) ocean forecasting system, focusing on the northwestern Pacific region. Results demonstrate significant improvements in forecast accuracy, with Root Mean Square Error (RMSE) reductions ranging from 38.8% for day 2 forecasts to 17.6% for day 5 forecasts. Spatial analysis reveals the method’s robust performance across diverse oceanographic regimes, including complex coastal and shelf regions where traditional models often struggle. The Transformer model showed the ability to capture and reproduce error patterns, effectively addressing both large-scale systematic biases and smaller-scale regional variations. The consistent performance across different forecast horizons suggests potential for extending the reliable forecast range of SST predictions. The findings have important implications for applications requiring precise SST forecasts, including operational oceanography, marine weather forecasting, and coupled ocean-atmosphere modeling.