Eddy Suprihadi

This study examines whether higher temporal resolution and more flexible model specification improve short-horizon forecasting of aggregate U.S. listed-real-estate exchange-traded fund returns. Using release-aware daily, monthly, and quarterly predictors, we evaluate VNQ as the primary target and IYR as a closely related robustness proxy over one-, two-, and four-week horizons, with a fixed 2022–2025 out-of-sample period. Seven forecasting systems are compared, including a Historical Mean benchmark, ARX-Ridge, Weekly LSTM, MIDAS-ADL, Mixed-Frequency VAR, Frequency-Specific LSTM, and Cross-Frequency Attention. Weekly-versus-native comparisons are interpreted as system-level contrasts, while a separate matched daily predictor experiment provides a narrower sensitivity check on sampling resolution. The Historical Mean achieves the lowest RMSE in all six target–horizon cases. The system-level comparisons show no consistent advantage for native-frequency implementations, and the matched experiment likewise provides no general evidence that retaining daily resolution improves forecasts. MIDAS-ADL exhibits specification-specific long-horizon instability associated with amplified fitted contributions, particularly from the monthly block, although annual re-estimation substantially attenuates this behavior. Forecast performance, therefore, depends jointly on temporal representation, model specification, information integration, and updating policy.

Previous studies on regional construction cost estimation in Indonesia have predominantly relied on uniform national indices or conventional econometric models, which often fail to capture spatial heterogeneity and evolving temporal trends. This study introduces a data-driven hybrid clustering framework that integrates KMeans and HDBSCAN to classify Indonesian districts and cities using Construction Cost Index (CCI) dynamics from 2014 to 2024, with economic plausibility assessed using Gross Regional Domestic Product (GRDP). Empirically, KMeans yields three interpretable macro-clusters, but produces a highly imbalanced partition in which the dominant cluster contains 474 districts (over 90% of observations), indicating limited sensitivity to localized irregularities. In contrast, HDBSCAN identifies 35 micro-clusters and flags 103 districts as noise, explicitly isolating anomalous trajectories that centroid-based clustering may absorb. Cluster validity is supported by statistically significant between-cluster differences in CCI (one-way ANOVA on 2020 CCI: KMeans F = 1696.49, p = 2.35 × 10⁻²²⁶; HDBSCAN F = 370.91, p = 9.22 × 10⁻²⁶⁷, excluding noise). The temporal profiles reveal heterogeneous cost regimes, including persistently high-cost zones, transitional regions exhibiting structural correction, and stable low-cost areas. By exposing these hidden disparities and providing an anomaly-sensitive zoning structure, the proposed framework supports differentiated budgeting, targeted subsidies, and adaptive infrastructure planning for geographically diverse settings such as Indonesia.

Forecasting returns in Real Estate Investment Trust (REIT) markets remains challenging because REIT performance is shaped by nonlinear and time-varying interactions with macro-financial conditions. This study evaluates the forecasting performance of Long Short-Term Memory (LSTM) neural networks relative to a TVP–VAR proxy implemented as an expanding window VAR for weekly U.S. U.S. REIT returns. All models are assessed within a harmonized experimental framework that applies consistent data preprocessing, feature construction, and strictly time-ordered out-of-sample evaluation. The results indicate that the baseline LSTM model delivers modest but more stable error-based performance than the TVP–VAR proxy, with improvements concentrated in RMSE and MAE, while evidence for directional predictability is weak and not consistently distinguishable from benchmark performance. To enhance transparency, SHapley Additive exPlanations (SHAPs) are used to interpret the LSTM forecasts. The attribution analysis highlights recent REIT returns, global equity indicators—particularly the Hang Seng Index—and crude oil prices as influential predictors, and shows that their contributions vary across volatility regimes, consistent with time-varying spillovers and changing risk transmission. Overall, the study positions LSTM forecasting combined with SHAP-based interpretation as a transparent and reproducible framework for comparative evaluation and driver analysis in weekly REIT returns, rather than as a strong directional timing tool.

This study investigates regime-dependent forecasting of the Saudi stock market by combining macro-controlled dependence analysis with nonlinear predictive modeling. Using daily data from September 2010 to August 2025, we analyze the interaction between the Tadawul All Share Index (TASI) returns and crude oil returns while controlling for inflation and interest-rate dynamics. A four-variable VAR with macro controls is estimated separately in pre- and post-COVID regimes to characterize directional predictability and changes in transmission lags. We then evaluate out-of-sample return forecasting performance across econometric benchmarks (ARIMA, ARIMAX, and VAR) and machine learning models (LSTM and XGBoost) under a strictly time-ordered expanding-window design with sequential train/validation/test partitioning. The results indicate that traditional linear benchmarks exhibit limited predictive ability in both regimes, with negative out-of-sample explanatory power. By contrast, XGBoost delivers the strongest overall performance, achieving positive out-of-sample R² in both regimes (0.046 in pre-COVID and 0.010 in post-COVID), together with the lowest forecast errors (RMSE = 0.0081 pre-COVID; 0.0078 post-COVID). Interpretability analysis further reveals a regime-sensitive shift in drivers: short-horizon equity lag dynamics dominate during stable periods, whereas oil-related and macro-financial variables gain importance under turbulent conditions. Economic-value evaluation supports the practical relevance of these gains, showing that XGBoost-based signals yield superior risk-adjusted trading outcomes and remain favorable under downside-risk and drawdown-based assessment. Overall, these findings highlight that forecasting in oil-linked emerging markets is inherently regime-dependent and that nonlinear ensemble learners, particularly XGBoost, provide a more robust and economically meaningful approach under structural change.

Forecasting stock market crashes is difficult because such events are rare, highly nonlinear, and shaped by latent structural and behavioral forces. This study introduces a calibrated and interpretable Random Forest framework for detecting pre-crash conditions through structural feature engineering, early-warning calibration, and model explainability. Using daily data on global equity indices and major large-cap stocks from the U.S., Europe, and Asia, we construct a feature set that captures volatility expansion, moving-average deterioration, Bollinger Band width, and short-horizon return dynamics. Probability-threshold optimization significantly improves sensitivity to rare events and yields an operating point at a crash-probability threshold of 0.33. Compared with econometric and machine learning benchmarks, the calibrated model attains higher precision while maintaining competitive F1 and MCC scores, and it delivers meaningful early-warning signals with an average lead-time of around 60 days. SHAP analysis indicates that predictions are anchored in theoretically consistent indicators, particularly volatility clustering and weakening trends, while robustness checks show resilience to noise, structural perturbations, and simulated flash crashes. Taken together, these results provide a transparent and reproducible blueprint for building operational early-warning systems in financial markets.

This study investigates the influence of input window size and ensemble size on the forecasting performance of Decision Tree-based ensemble methods in one-step-ahead financial return prediction. Nineteen financial instruments spanning stocks, indices, commodities, and cryptocurrencies are evaluated under a chronological train-validation-test framework. Daily closing prices are transformed into log returns, and models are trained across rolling windows of 5, 10, 20, and 60 observations. The methods compared include Naive forecasting, Linear Regression, a standalone Decision Tree, Bagging-DT, Boosting-DT, and Stacking-DT, with ensemble configurations selected based on validation MAE. Linear Regression achieves the lowest average magnitude error and the best overall rank, while Bagging-DT and Boosting-DT substantially outperform both Naive forecasting and the standalone Decision Tree. The validation-selected ensemble reduces MAE across all 19 instruments, yielding an average out-of-sample improvement of 31.89 % relative to the single tree. A test-best diagnostic comparison suggests a higher improvement ceiling of 35.10 %, though this figure serves solely as an upper-bound reference. A 20-day window delivers the strongest average error performance, while a 60-day window and 100-estimator ensembles are most frequently selected at the individual instrument level. Larger ensembles incur additional computational cost without producing consistent accuracy gains. Friedman and Wilcoxon tests confirm significant performance differences among models, and Diebold-Mariano tests indicate that ensemble methods consistently outperform both Naive forecasting and the standalone Decision Tree. Overall, Decision Tree ensembles offer robust improvements over persistence-based and single-tree forecasting, though their advantage relative to Linear Regression remains instrument-dependent.