Climate modeling for South Asia: statistical and deep learning for rainfall and temperature prediction – Nature

Executive Summary

This report presents a comparative analysis of four hydrometeorological forecasting models—SARIMA, TDNN, LSTM, and XGBoost—to enhance climate resilience and support the achievement of Sustainable Development Goals (SDGs) in South Asia. Utilizing a centennial dataset (1901–2023) of monthly rainfall and temperature for seven countries, this study establishes a reproducible pipeline for model evaluation. The framework’s novelty lies in its cross-country benchmarking of classical and deep learning methods under standardized preprocessing and cross-validation. Findings indicate that model performance is contingent on the specific climatic variable and geographical region. For instance, TDNN and XGBoost are better suited for nonlinear rainfall dynamics, directly impacting SDG 2 (Zero Hunger) and SDG 6 (Clean Water and Sanitation). LSTM models excel at forecasting minimum temperatures, which is critical for agricultural planning and public health warnings under SDG 11 (Sustainable Cities and Communities). Conversely, SARIMA remains effective for stable, seasonal series. These insights are consolidated into a hybrid model selection guide to aid decision-making. The resulting 2024–2025 forecasts, complete with uncertainty bands, provide actionable intelligence for policies related to crop calendars, irrigation, flood preparedness, and water resource management, thereby advancing SDG 13 (Climate Action) across the region.

1.0 Introduction: Aligning Climate Forecasting with Sustainable Development Goals

1.1 Context: Climate Vulnerability in South Asia and the Urgency for Action

South Asia’s susceptibility to monsoonal variations and extreme weather events presents a significant obstacle to sustainable development. Reliable hydrometeorological forecasting is essential for mitigating risks and managing resources, directly contributing to several SDGs. The region, encompassing Afghanistan, Bangladesh, Bhutan, India, Sri Lanka, Nepal, and Pakistan, faces recurrent climate-related hazards such as severe monsoons, heatwaves, and droughts. These events threaten food security, water availability, and community safety, undermining progress toward key development targets.

Accurate monthly forecasts of rainfall and temperature are critical for climate adaptation strategies. Such forecasts enable proactive measures that support:

• SDG 2 (Zero Hunger): by informing crop calendars and irrigation schedules to stabilize agricultural yields.
• SDG 6 (Clean Water and Sanitation): by aiding in the sustainable management of water resources and allocation.
• SDG 11 (Sustainable Cities and Communities): by strengthening disaster preparedness for floods and heatwaves.
• SDG 13 (Climate Action): by building resilience and adaptive capacity to climate-related hazards.

While various forecasting methodologies exist, a systematic, multi-country comparison over a centennial timeframe has been lacking. This study addresses this gap by providing a robust framework for selecting the most effective forecasting tools to support climate-informed decision-making aligned with the SDGs.

1.2 Research Objectives and Contribution to SDGs

This research aims to benchmark classical, deep learning, and machine learning models to identify the most effective approaches for hydrometeorological forecasting in South Asia. The primary objectives are:

1. To develop a unified forecasting framework that integrates statistical (SARIMA) and machine learning (TDNN, LSTM, XGBoost) models.
2. To evaluate the predictive accuracy of each model across diverse climatic variables and countries using standardized performance metrics.
3. To formulate a practical model-selection guide that maps climatic conditions to the most suitable forecasting model, thereby providing a robust tool for regional climate action.

By achieving these objectives, this study contributes directly to evidence-based policymaking for climate resilience, providing a methodological foundation for advancing the 2030 Agenda for Sustainable Development in one of the world’s most climate-vulnerable regions.

2.0 Methodological Framework for Climate Resilience

2.1 Data and Study Area

The analysis is based on a comprehensive monthly dataset from the World Bank Group’s Climate Change Knowledge Portal, covering the period from January 1901 to December 2023. The dataset includes rainfall, minimum temperature, and maximum temperature for seven South Asian nations. The extensive temporal scope of 1,476 observations per series ensures that the models are trained and validated against a wide range of climatic variability, enhancing their reliability for long-term strategic planning related to climate action (SDG 13).

2.2 Forecasting Models for Sustainable Planning

A unified pipeline was developed in Python to benchmark four distinct forecasting models, each chosen for its unique strengths in handling time-series data:

• SARIMA (Seasonal Autoregressive Integrated Moving Average): A classical statistical model effective for data with clear and stable seasonal patterns.
• TDNN (Time Delay Neural Network): A feedforward neural network capable of capturing short-term temporal dependencies and nonlinear patterns, crucial for predicting erratic rainfall.
• LSTM (Long Short-Term Memory): A recurrent neural network designed to model long-range dependencies, making it suitable for temperature series with long memory.
• XGBoost (Extreme Gradient Boosting): A powerful machine learning algorithm that excels in handling complex interactions and noisy data through an ensemble of decision trees.

Standardized preprocessing, including normalization and feature engineering, was applied to ensure a fair and direct comparison across all models.

2.3 Evaluation and Validation Protocol

A rigorous evaluation protocol was implemented to ensure the robustness and generalizability of the findings. Key components of this protocol include:

• Cross-Validation: A 5-fold expanding TimeSeriesSplit was used to assess out-of-sample performance, ensuring that the models are validated on unseen data.
• Hyperparameter Optimization: A randomized search was conducted to tune the hyperparameters for TDNN, LSTM, and XGBoost models, optimizing their predictive accuracy.
• Standardized Metrics: Model performance was compared using a comprehensive set of metrics: Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Coefficient of Determination (R²), and Cross-Validated RMSE (CV-RMSE).

This systematic approach ensures that the selected models provide reliable and accurate forecasts, which are fundamental for effective planning and action toward achieving climate-related SDGs.

3.0 Results: Model Performance and Implications for SDG Targets

3.1 Comparative Model Evaluation

The comparative analysis revealed that no single model is universally superior; performance is highly dependent on the specific characteristics of the climate variable and region. This finding underscores the need for a tailored approach to forecasting to support specific SDG-related interventions.

• Rainfall Forecasting (SDG 2 & SDG 6): TDNN and XGBoost models generally outperformed others in forecasting rainfall, which is characterized by nonlinear dynamics. For example, XGBoost achieved an R² of 0.97 for India’s rainfall, demonstrating high accuracy crucial for agricultural planning and water management.
• Minimum Temperature Forecasting (SDG 2 & SDG 11): LSTM models proved most effective for minimum temperature series, which exhibit long-term memory. In Afghanistan, LSTM achieved an R² of 0.95 with a low RMSE, providing reliable data for predicting cold surges that affect crops and human settlements.
• Stable Climate Patterns: SARIMA models performed well for series with regular and stable seasonality, such as the minimum temperature in Bangladesh and Sri Lanka. This makes it a reliable tool for contexts where climate patterns are historically consistent.

3.2 A Hybrid Model Selection Guide for Targeted Climate Action

Based on the performance evaluation, a hybrid model selection guide was developed. This guide recommends the optimal forecasting model based on the variable and location, enabling policymakers to deploy the most accurate tools for their specific needs. This targeted approach enhances the effectiveness of climate adaptation strategies, ensuring that resources are allocated based on robust evidence. For instance, a water management authority in India could use XGBoost for rainfall forecasts (supporting SDG 6), while an agricultural agency in Nepal could use LSTM for temperature forecasts to protect crops (supporting SDG 2).

3.3 Forecasts for 2024–2025: Informing SDG-Related Policies

Using the best-performing model for each variable, forecasts for 2024–2025 were generated. These forecasts provide actionable, forward-looking information for short-term and medium-term planning. The inclusion of uncertainty bands allows for risk-informed decision-making, which is critical for building resilience. These outputs can be directly integrated into:

1. Early Warning Systems: For floods and heatwaves, contributing to SDG 11.
2. Resource Allocation Plans: For water and agricultural inputs, advancing SDG 6 and SDG 2.
3. Climate Adaptation Roadmaps: Providing quantitative evidence to support national and regional strategies under SDG 13.

4.0 Conclusion and Recommendations for Policy Integration

4.1 Key Findings and Methodological Contributions

This study successfully benchmarked four leading forecasting models on a centennial hydrometeorological dataset for South Asia. The key conclusion is that a hybrid, data-driven approach is superior to a one-size-fits-all strategy. The research contributes a reproducible and transparent pipeline for model selection and evaluation, providing a valuable resource for both researchers and practitioners working on climate adaptation. The findings confirm that advanced machine learning models like LSTM and XGBoost offer significant advantages for capturing the complex and nonlinear dynamics of the region’s climate, which is essential for building resilient systems.

4.2 Policy Recommendations for Achieving Climate-Related SDGs

To translate these findings into tangible progress on the SDGs, the following recommendations are proposed:

• Integrate Hybrid Forecasting into National Planning: Governments and regional bodies should adopt the proposed hybrid model-selection framework to enhance the accuracy of their climate information systems. This will improve planning for agriculture (SDG 2), water management (SDG 6), and disaster risk reduction (SDG 11).
• Develop Sector-Specific Early Warning Systems: Use the generated forecasts to create tailored early warning systems for farmers, water managers, and urban planners. This proactive approach supports climate action (SDG 13) by enabling timely interventions.
• Invest in Data Infrastructure and Capacity Building: To ensure the sustainability of such data-driven approaches, investment in high-quality data collection and local analytical capacity is crucial.

4.3 Limitations and Future Research Directions

The study’s limitations include its reliance on univariate models, which do not account for interactions between different climate variables or large-scale drivers like ENSO. Future research should explore multivariate and ensemble models to capture these complex dynamics. Additionally, incorporating exogenous predictors could further enhance forecast accuracy. The ultimate goal should be to link these advanced predictive outputs directly with sectoral decision-support tools, such as crop yield models and water allocation systems, to fully operationalize climate science for sustainable development.

Analysis of Sustainable Development Goals (SDGs) in the Article

1. Which SDGs are addressed or connected to the issues highlighted in the article?

• SDG 2: Zero Hunger: The article directly connects hydrometeorological forecasting to agriculture. It states that the forecasts are intended to “inform crop calendars” and “irrigation scheduling.” Accurate predictions of rainfall and temperature are essential for optimizing agricultural practices, ensuring food security, and building resilient farming systems, which are central to achieving Zero Hunger.
• SDG 6: Clean Water and Sanitation: The research aims to improve “water allocation” and “water management.” By providing reliable forecasts of rainfall, the models discussed in the article can help authorities manage water resources more effectively, which is a key component of SDG 6, particularly in a region like South Asia where water availability is heavily influenced by monsoons.
• SDG 11: Sustainable Cities and Communities: The article highlights the importance of forecasting for “flood readiness” and issuing “heat warnings.” These applications are crucial for disaster preparedness and risk reduction in human settlements, aligning with the goal of making cities and communities more resilient to climate-related hazards.
• SDG 13: Climate Action: This is the most prominent SDG addressed. The entire study is focused on improving the prediction of climate-related events like “severe monsoons, heatwaves, droughts, and cold surges.” The development of advanced forecasting models is a direct contribution to strengthening resilience and adaptive capacity to climate-related hazards and supports the creation of early warning systems. The article explicitly mentions that its goal is to contribute to “climate adaptation” and “climate-informed decision-making.”

2. What specific targets under those SDGs can be identified based on the article’s content?

• Target 2.4: “By 2030, ensure sustainable food production systems and implement resilient agricultural practices that increase productivity and production, that help maintain ecosystems, that strengthen capacity for adaptation to climate change, extreme weather, drought, flooding and other disasters…” The article’s framework, which produces forecasts to inform “crop calendars” and “irrigation scheduling,” directly supports the implementation of resilient agricultural practices that can adapt to the climatic variability and extreme weather events discussed.
• Target 6.4: “By 2030, substantially increase water-use efficiency across all sectors and ensure sustainable withdrawals and supply of freshwater to address water scarcity…” The forecasts for rainfall patterns are intended to inform “water allocation,” which is essential for managing freshwater supplies efficiently and sustainably, especially in the face of unpredictable monsoons.
• Target 11.5: “By 2030, significantly reduce the number of deaths and the number of people affected and substantially decrease the direct economic losses relative to global gross domestic product caused by disasters, including water-related disasters…” The article’s emphasis on improving forecasts for “flood readiness” and “heat warnings” directly contributes to disaster risk reduction, which is the core of this target. Better early warnings can help protect vulnerable populations from extreme weather events.
• Target 13.1: “Strengthen resilience and adaptive capacity to climate-related hazards and natural disasters in all countries.” The central theme of the article is to enhance forecasting of hydrometeorological events (“monsoons, heatwaves, floods”) in South Asia. This work directly strengthens the region’s adaptive capacity by providing tools for better planning and response to climate-related hazards.
• Target 13.3: “Improve education, awareness-raising and human and institutional capacity on climate change mitigation, adaptation, impact reduction and early warning.” The study’s development of a “reproducible pipeline,” a “hybrid selection guide,” and a systematic benchmarking of different models serves to build institutional capacity for climate forecasting. This contributes to improving early warning systems and climate-informed decision-making, as mentioned in the conclusion.

3. Are there any indicators mentioned or implied in the article that can be used to measure progress towards the identified targets?

• Climatic Data Variables: The core data used in the study—”monthly rainfall,” “minimum temperature,” and “maximum temperature”—are themselves fundamental indicators used to monitor climate patterns and variability. These are direct inputs for tracking climate-related hazards.
• Forecasting Model Performance Metrics: The article explicitly mentions several metrics used to evaluate the accuracy of its forecasting models: “RMSE, MAPE, R², and CV-RMSE.” The improvement of these metrics can be seen as a proxy indicator for the enhanced reliability of early warning systems. A lower error rate (RMSE, MAPE) and higher coefficient of determination (R²) signify a more effective forecasting tool for climate adaptation.
• Development of Forecasting Tools: The creation and application of the “unified, reproducible pipeline” and the “practical model-selection guide” are indicators of increased institutional capacity (Target 13.3). The existence and use of such advanced, data-driven frameworks demonstrate progress in a country’s ability to manage climate risks.
• Generation of Forecasts: The article’s output of “2024–2025 forecasts with uncertainty bands” is a tangible indicator. The availability and dissemination of such forecasts to inform “crop calendars, irrigation scheduling, flood preparedness, heat warnings, and water allocation” directly measure the operationalization of early warning and climate adaptation strategies.

4. Summary Table of SDGs, Targets, and Indicators

Source: nature.com