Modelling in Complex Systems

Saving Composition, Digital Payments and Household Financial Resilience in Developing Economies

Dr Nur Syazwani Mazlan

We examine how the composition of a household’s financial inclusion, rather than its level, relates to its capacity to raise emergency funds. Using data from more than 300,000 respondents across 77 developing and emerging economies and four waves of the Global Findex survey, we differentiate between saving at formal financial institutions and saving through informal savings groups. Both forms of saving are positively associated with the ability to raise emergency funds, although the association is stronger for formal saving. When households have access to both channels, the incremental benefit of each is reduced, indicating that formal and informal saving act as partial substitutes rather than complements. Digital bill payments are also positively associated with financial resilience and similarly substitute for, rather than reinforce, saving-based mechanisms. Macroeconomic conditions are associated with resilience in ways consistent with consumption-smoothing theory. We further show that common measures of financial inclusion place at least as much weight on informal as on formal saving, implying that improvements in measured inclusion can mask reductions in financial resilience if formal saving replaces informal saving. Finally, we reject the conventional saving-versus-borrowing factor structure: financial inclusion behaviors are better characterized by their degree of formality, and imposing the conventional structure obscures these characteristics. Estimates throughout are conditional associations.

Agricultural Production and Greenhouse Gas Emission: Dynamic Stochastic Frontier Model with Extreme Weather Indicators

Dr Lee How Chinh

Sustaining agricultural production while managing greenhouse gas emissions requires an understanding of how inputs, weather conditions and institutional factors relate to both outcomes. This presentation examines production and emissions performance across overall agriculture, livestock and crops using unbalanced annual panels covering 147, 151 and 47 reporting areas, respectively, with observations spanning 2003–2023 for overall agriculture and livestock and 2005–2023 for crops.

Motivated by stochastic frontier analysis, the study develops an iterative two-stage framework linking dynamic production and emissions benchmarks. The first stage incorporates current inputs, lagged inputs and previous-year outcomes. The second stage uses linked generalised additive models to relate residual-based performance indices to institutional factors and monthly precipitation and average maximum temperature, allowing nonlinear weather associations.

The results reveal strong persistence in production and emissions, while input and institutional associations vary across agricultural systems. Relationships between the fitted performance indices are negative for overall agriculture and crops but positive for livestock. Geographic comparisons further show that areas with high relative production performance do not necessarily exhibit high relative emissions performance.

These findings demonstrate the value of assessing production and emissions jointly. The presentation also discusses methodological limitations, particularly the interpretation of the indices as sample-relative performance measures, and priorities for validating the proposed framework.

Climate-Production Capacity and Environmental Disclosure in Chinese Agricultural Firms: Frontier Shifts and Efficiency Gaps

Shijie Wang

We examine how climate protection capacity and environmental disclosure relate to the joint economic and environmental performance of Chinese agricultural firms. We distinguish between the potential performance frontier and how closely firms operate to that frontier. We develop a two-stage iterative carbon productivity frontier model that incorporates firms’ production inputs, environmental disclosure, and exposure to provincial climate protection conditions. We find that stronger climate protection capacity is associated with a higher attainable carbon productivity frontier. We also find that more extensive environmental disclosure is associated with lower technical inefficiency and greater proximity to the frontier. We further document variation in estimated efficiency across disclosure intensity, firm characteristics, and provincial policy environments. We highlight the complementary roles of external climate protection conditions and internal environmental governance. Our findings suggest that public support may expand firms’ performance opportunities, while effective environmental management and policy implementation may help firms realize those opportunities.

From Climate Risk to Food Security: Policy Pathways through Agricultural Productivity and Water Management

Dr Hashem Salarzadeh Jenatabadi

Climate change poses increasing challenges to agricultural productivity and global food security, particularly through rising temperatures, changing precipitation patterns, and growing pressure on freshwater resources. This study investigates the relationships between climate change, agricultural total factor productivity (TFP), and multidimensional food security using a longitudinal panel data framework. The analysis addresses four interconnected research questions across three analytical phases. First, it examines whether agricultural TFP mediates the relationships between temperature, rainfall, their interaction, and overall food security. Second, it investigates freshwater resources as a mediating mechanism linking climatic conditions and carbon emissions to agricultural productivity, while assessing their moderating role in fertilizer and pesticide effectiveness. Third, it explores sequential relationships among food availability, accessibility, and utilization. Finally, it evaluates how temperature–rainfall interactions influence food security through these interconnected dimensions. The proposed analytical approach integrates panel econometric modelling, within–between decomposition, mediation analysis, and moderation analysis to capture complex relationships and country-level heterogeneity. By integrating agricultural production efficiency, environmental constraints, and multidimensional food security, the study aims to advance understanding of climate-related vulnerabilities and identify potential adaptation pathways. Its findings are intended to inform agricultural resource management, climate adaptation strategies, and evidence-based food security policies across diverse national contexts.

Keywords: Climate Change; Agricultural Total Factor Productivity; Food Security; Water Resources; Adaptation Policy; Panel Data

External Validation of Vision Foundation Model Representations for Malay Heritage Textiles: A Label-Grounded Evaluation of CLIP, DINOv2 and SigLIP 2

Dr Nazirul Hazim A. Khalim

Pre-trained vision models are widely used to organise digitised heritage collections, on the assumption that their image representations encode attributes a curator would recognise. The assumption is tested here on Malay heritage textiles, against labels stated by sources other than the models. The corpus comprises 775 images from four sources (a printed museum catalogue of limar, a weft-ikat silk, and the captioned photographs of two specialist collectors) and 144 crops of catalogue plates labelled by structural section. Six models are evaluated: CLIP ViT-B/32 and ViT-L/14, SigLIP 2, DINOv2 ViT-B/14 and ViT-L/14, and an ImageNet-supervised ViT-B/16. Each is given two inputs, the standard single view of 224 pixels and tiles cut at higher resolution. Every estimate is reported beside a baseline, with related images held out and intervals resampled over textiles or posts.

The source collection of an image is identified from its nearest neighbours in 77 to 94% of cases against a baseline of 35%, and in 73% from colour and texture statistics alone. Structural section is identified better from the single view (84 to 91% against 50%), and technique better from tiles (75 to 83% against 33%, and 83 to 90% with a linear classifier). A difference between model families in matching catalogue plates to photographs from another source is present under the single view and absent under tiles. Zero-shot labels depend on prompt wording: the share of limar plates labelled ikat ranges from 2% to 91% across 21 combinations of model, prompt and input. Place of origin could not be separated from technique.

We argue that source is the attribute these representations yield most consistently, that a claim about what a model encodes holds only for the input scale tested, and that clusters of heritage images cannot be interpreted without labels from outside the model.

Keywords: cultural heritage; textiles; vision foundation models; dataset bias; zero-shot classification; external validation

What Shapes Well-Being Over Time? Uncovering the Drivers of Quality of Life and Mental Health Through Longitudinal Modelling and Explainable AI

Dr Mogana Darshini Ganggayah and Dr Shin Zhu Sim

Authors: Mogana Darshini Ganggayah^1*, Shin Zhu Sim¹, Hashem Salarzadeh Jenatabadi¹, Resha Devi M. Vadivel¹, Lim Fang Yee¹, Lee Wai Leng², Erniel Bayhon Barrios¹, Abdul Rahman Mohamad Rom¹, Lee How Chinh¹, Nazirul Hazim A. Khalim¹, Nur Syazwani^1
1 School of Business, Monash University Malaysia. ² School of Science, Monash University Malaysia.

Background. Mental health and quality of life are influenced by complex interactions among demographic, socioeconomic, health, lifestyle, and behavioural factors. Understanding these relationships over time is essential for identifying vulnerable populations and developing targeted interventions. This study aims to investigate the determinants and predict mental health and quality of life among residents of the South East Asia Community Observatory (SEACO), Malaysia, using longitudinal data from 2013 and 2018.

Methods. Longitudinal data comprising 17,188 observations from 8,594 residents across two survey waves were analysed. Four outcomes were examined: depression, anxiety, stress, and overall quality of life. Longitudinal mixed-effects models were employed to identify factors associated with each outcome while accounting for repeated observations. Four machine learning algorithms, Elastic Net, Support Vector Regression, Random Forest, and XGBoost, were developed and compared using mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²). Elastic Net-based permutation feature importance was used to identify and rank the top 20 predictors for each outcome across both survey years. Explainable Artificial Intelligence (XAI), incorporating SHapley Additive exPlanations (SHAP), was applied to quantify predictor contributions and examine their direction of influence.

Contributions. The integrated framework will provide insights into longitudinal determinants, predictive performance, and the relative importance of factors influencing mental health and quality of life. Findings are expected to support evidence-informed interventions, early risk identification, and community-level health planning.

Keywords: Mental health; Quality of life; Longitudinal analysis; Machine learning; Explainable Artificial Intelligence; SEACO

Volatility Clustering in Cortical Dynamics: Conditional Heteroskedasticity as a Group Marker in the Alcoholism EEG Cohort

Dr Nazirul Hazim A. Khalim and Prof. Erniel Bayhon Barrios

Group comparisons of electroencephalography rest almost entirely on three feature families: spectral power, event-related potential components, and inter-channel connectivity. All three summarise the level or the covariation of the signal. This paper measures a fourth property, the persistence of the signal’s conditional variance — the degree to which a large voltage deviation makes the next one likely to be large, known in econometrics as volatility clustering — and asks whether it differs between alcohol-dependent and control subjects. Each single-trial channel recording is modelled as an AR(1) process with cluster-specific ARCH(1) errors, the ARCH slope α is taken as the estimand, and 111 pre-registered group contrasts are tested over the full UCI alcoholism EEG cohort — 122 subjects, 10,962 trials, 179.6 million samples — by constrained parametric bootstrap with joint Benjamini–Hochberg correction. Twenty-one of 111 contrasts reject at a false-discovery rate of 0.05, and every rejection has the same sign: alcohol-dependent subjects show lower volatility clustering than controls. The effect is anatomically concentrated. Only prefrontal cortex clears correction at whole-region granularity (α̂ = 0.428 against 0.981), the significant cells lie in the left prefrontal pool across all three stimulus conditions, and every interaction family is null at full panel scale. Because 19 of 86 control-side estimates sit on the estimator’s 0.99 bound, the reported gaps are lower bounds rather than point estimates: the direction is established, the magnitude is not. The ARCH slope is invariant to amplitude rescaling and measures temporal dependence in the conditional variance rather than its level, which distinguishes it from the amplitude-separable measures now in wide clinical use, the aperiodic spectral exponent chief among them. Whether it is empirically separable from those measures is left open and stated as the first item of further work.

Keywords: electroencephalography; volatility clustering; conditional heteroskedasticity; ARCH models; alcohol use disorder; constrained bootstrap