Research Article
Coverage and Protocol-Aware Evaluation of Risk Enrichment and Graph-Tabular Learning for Anti-Money Laundering and Fraud Detection
Issue:
Volume 12, Issue 5, October 2026
Pages:
97-113
Received:
13 July 2026
Accepted:
29 July 2026
Published:
8 September 2026
Abstract: Reported gains from graph-based machine learning in anti-money laundering and fraud detection can vary substantially with the historical evidence available to a model and with the evaluation protocol used. This study introduces a coverage-aware and protocol-aware evaluation framework for assessing account-risk enrichment, graph neural networks, tabular gradient boosting, and graph-tabular stacking across synthetic banking anti-money-laundering data, a real Bitcoin transaction graph, and a complementary real credit-card fraud dataset. The analysis compares account-risk enrichment, four topology-only graph neural networks, XGBoost, LightGBM, and embedding-based stacking under warm-start, account-grouped, cold-start, coverage-sensitivity, and temporal evaluation settings. On the synthetic HI-Small benchmark, strict account grouping removes source-account historical coverage by construction and substantially limits the opportunity for enrichment to contribute useful signal. Under the warm-start protocol, enrichment produces only a small change in discrimination, while the coverage-sensitivity analysis shows no reliable improvement in ROC-AUC as historical coverage increases. PR-AUC results instead indicate a small but consistent performance cost at most tested coverage levels after correction for multiple comparisons. On the lower-coverage LI-Small benchmark, enrichment again shows a small negative effect that does not remain significant after manuscript-wide correction. External validation on the Elliptic Bitcoin graph shows that topology-only graph neural networks do not automatically outperform strong tabular models. GraphSAGE is the strongest graph neural network tested, but XGBoost and LightGBM achieve clearly higher ROC-AUC and PR-AUC. Graph-tabular stacking is also dataset dependent: it yields small positive gains on Elliptic, while tabular features alone are competitive with or superior to stacking on HI-Small. These results show that graph-based improvements should not be interpreted independently of historical coverage, split construction, and feature redundancy. Practical evaluation should report coverage and protocol alongside graph-derived claims, use strong tabular baselines, and complement ROC-AUC with PR-AUC and calibration-oriented metrics before concluding that a graph-based method provides a robust advantage.
Abstract: Reported gains from graph-based machine learning in anti-money laundering and fraud detection can vary substantially with the historical evidence available to a model and with the evaluation protocol used. This study introduces a coverage-aware and protocol-aware evaluation framework for assessing account-risk enrichment, graph neural networks, tab...
Show More
Research Article
Comparative Study of LSTM, XGBoost and Hybrid LSTM-XGBoost Models for Rainfall Forecasting in Semi-arid Regions
Rony Mutugi Muriithi
,
Peter Kinyua Gachoki,
Mutua Kilai*
Issue:
Volume 12, Issue 5, October 2026
Pages:
114-128
Received:
11 July 2026
Accepted:
28 July 2026
Published:
8 September 2026
DOI:
10.11648/j.ijdsa.20261205.12
Downloads:
Views:
Abstract: Rainfall forecasting remains challenging in semi-arid regions due to high variability and intermittent rainfall patterns. Statistical forecasting methods such as Autoregressive Integrated Moving Average(ARIMA) and Seasonal Autoregressive Integrated Moving Average (SARIMA) often struggle to capture the non-linear dynamics typical of such rainfall. Machine learning (ML) techniques such as Extreme Gradient Boosting (XGBoost) and Long Short-Term Memory (LSTM) networks offer improved forecasting performance but have individual limitations. This study compares LSTM, XGBoost and a hybrid LSTM-XGBoost model for monthly rainfall forecasting, using SARIMA as a baseline. The study utilizes 44 years of CHIRPS (Climate Hazards Group InfraRed Precipitation with Station) data from Tana River County, Kenya, accessed through Google Earth Engine. Data preprocessing included log transformation, stationarity testing, normalization, and feature engineering. The data was chronologically split to form the train, validation and test sets. The performance of the models was evaluated using Root Mean Square Error, Mean Absolute Error, Coefficient of Determination and Nash Sutcliffe Efficiency. Shapley Additive Explanations were used for interpretability. The XGBoost, LSTM and Hybrid LSTM-XGBoost models outperformed the baseline, achieving a Coefficient of Determination of approximately 0.61 compared with 0.52. The hybrid model performed best overall, with an RMSE of 38.68 mm, MAE of 17.23 mm, and R2 of 0.609. Hybrid forecasting approaches combining deep learning and machine learning should be further explored for rainfall prediction in semi-arid regions due to their potential to capture complex temporal patterns.
Abstract: Rainfall forecasting remains challenging in semi-arid regions due to high variability and intermittent rainfall patterns. Statistical forecasting methods such as Autoregressive Integrated Moving Average(ARIMA) and Seasonal Autoregressive Integrated Moving Average (SARIMA) often struggle to capture the non-linear dynamics typical of such rainfall. M...
Show More
Research Article
The Impact of Varying Window Sizes on the Performance of a Hybrid CNN-LSTM Model A Case of Forecasting USD VS KES
Issue:
Volume 12, Issue 5, October 2026
Pages:
129-138
Received:
18 August 2026
Accepted:
4 September 2026
Published:
28 September 2026
Abstract: Financial time series are characterized by non-linearity, inherent volatility, and changing temporal patterns making accurate predictions very challenging. This study deploys a CNN-LSTM hybrid model to predict the USD/KES exchange rate giving particular emphasis to the effect the size of the historical input window has on the hybrid model forecasting performance. The study relied on Daily USD/KES exchange-rate data obtained from the Central Bank of Kenya to train, evaluate and validate the hybrid model. This data was prepossessed and transformed into sequential inputs using different window configurations, allowing the models to learn from varying amounts of historical information. The study combined a CNN-based local pattern extraction with an LSTM-based temporal dependency learning model to develop a hybrid CNN-LSTM model. The hybrid's model performance was then subjected to a comparison with standalone CNN and LSTM model and evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) The hybrid CNN-LSTM prediction results indicated that the model was more effective in prediction among the evaluated architecture. The choice of the input window sizes affected the model's ability to learn and capture trends in the time series data. The results from the rolling-window experiments demonstrated that shorter and moderately sized historical windows can provide a more responsive representation of recent exchange-rate movements than an expanding window, particularly during periods of pronounced market variation. These findings highlight the importance of considering both model architecture and input-window design when developing deep-learning approaches for USD/KES exchange-rate forecasting. The study provides an empirical basis for selecting appropriate historical window sizes when applying hybrid deep-learning models to volatile financial time series in the Kenyan foreign-exchange market.
Abstract: Financial time series are characterized by non-linearity, inherent volatility, and changing temporal patterns making accurate predictions very challenging. This study deploys a CNN-LSTM hybrid model to predict the USD/KES exchange rate giving particular emphasis to the effect the size of the historical input window has on the hybrid model forecasti...
Show More
