Gradient boosting models predict retail demand effectively, yet inventory planners distrust predictions they cannot interpret. This paper addresses the trust gap through a four-level Explainable AI (XAI) framework that pairs LightGBM prediction with layered interpretability: Permutation Feature Importance (PFI) for global ranking, Partial Dependence Plots (PDP) for marginal effects, SHapley Additive exPlanations (SHAP) for instance-level attribution, and Individual Conditional Expectation (ICE) plots for heterogeneity detection. ICE analysis, the key extension beyond prior work, reveals when averaged trends mask divergent product behavior, justifying segment-specific inventory policies. Validated on the UCI Online Retail II dataset with 41 engineered features, the proposed model achieves RMSE of 33.71 and R(2 )of 0.243 on inherently intermittent stock-keeping unit (SKU)-level daily demand, significantly outperforming most baselines (paired t-test, p <0.001; CatBoost not significant, p = 0.917). Ablation experiments confirm rolling statistics as the dominant feature group ( Delta R-2=-0.064 upon removal). ICE heterogeneity analysis identifies product clusters where temporal features produce opposite effects, demonstrating that uniform reorder policies systematically misallocate inventory for specific segments.
Modern software development relies on automated build systems that compile and test code whenever developers make changes. Predicting whether these builds will succeed or fail before execution could save computational resources and developer time. However, many machine learning models for build prediction suffer from temporal data leakage, a methodological flaw where the model inadvertently uses information that would only be available after the build completes, producing artificially inflated accuracy that fails in real-world deployment. This study develops a three-type taxonomy to systematically identify and prevent such leakage: (1) Direct Outcome Encoding (using the build result itself as a feature), (2) Execution-Dependent Metrics (information generated during build execution), and (3) Future Information Leakage (using data from chronologically later builds). Applying this taxonomy reveals that prior studies reporting 95-99% accuracy likely used contaminated features, while realistic accuracy is substantially lower. The methodology is validated on 175,706 builds from two open-source CI/CD platforms spanning 10 years: TravisTorrent (100,000 builds, 2013-2017) and GHALogs (75,706 workflows, 2023). Removing leaky features reduces accuracy by 15.07 percentage points on TravisTorrent (97.8% to 82.73%) but only 0.48 points on GHALogs (83.77% to 83.30%), revealing that modern GitHub Actions' tight integration with repositories enables accurate prediction from static project metadata alone. Using only legitimately available pre-build features, Random Forest classifiers achieve 82.73% (TravisTorrent) and 83.30% (GHALogs) accuracy, sufficient for practical deployment. Surprisingly, project maturity and build history prove more predictive than code complexity metrics, suggesting organizational factors outweigh code quality. The models generalize across programming languages (Java, Ruby, Python, JavaScript) with minimal performance variation. Open-source tools for detecting temporal leakage in any software prediction task are provided.
Retail marketing measurement increasingly requires granular campaign-level insights without relying on user-level tracking. However, the two dominant approaches, Marketing Mix Modeling (MMM) and Multi-Touch Attribution (MTA), often produce fragmented insights. MMM is privacy-safe and robust for channel-level planning but is too coarse for campaign optimization, while MTA provides granular attribution but has become less reliable under increasing privacy restrictions. We propose Integrated Marketing Attribution (IMA), a unified framework that combines MMM with channel specific Bayesian attribution models to derive campaign-level effects from aggregated data. By leveraging MMM-informed priors, IMA delivers granular, privacy-safe attribution while preserving consistency with MMM.
Due to the changing nature of online transactions and the growing sophistication of cybercriminals, the fast expansion of digital banking systems has greatly raised the demand for efficient fraud detection. More sophisticated systems must be developed since conventional fraud detection techniques frequently fail to meet these demands. In order to identify fraud in real-time financial transactions, this research suggests a hybrid deep learning model that blends Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks. LSTMs describe temporal dependencies, allowing the system to comprehend both the individual transaction characteristics and their sequences across time, whereas CNNs are utilized to extract spatial aspects from transaction data. The model is evaluated using publicly accessible financial transaction datasets and contrasted with conventional machine learning algorithms, such as support vector machines and decision trees. According to experimental data, the hybrid CNN-LSTM model performs better than traditional techniques in terms of computing efficiency, recall, and detection accuracy. The results show that the suggested model is a useful tool for digital banking systems since it is especially good at spotting fraudulent activity in real time. This strategy gives financial institutions a strong way to protect transactions in a world that is becoming more and more digital by offering notable advancements in the early detection of fraud.
DevOps toolchains record 164 metrics across eight software development lifecycle (SDLC) phases, yet no public dataset links these pre-deployment signals to production runtime outcomes. We present PRESTO, a framework for predicting software system performance from multi-phase SDLC metrics, plus a Gaussian copula-based synthetic data generator—both evaluated as a proof of concept, not a validated prediction method. Under nested cross-validation that selects the algorithm and hyperparameters within the training partition, non-autoregressive SDLC process features reach R2=0.108 to 0.620 across three synthetic domains; a target-correlation ablation shows that this is largely recovery of the generator’s specified structure, shared identically across all three domains, so the results show that the framework recovers relationships embedded in the generator. Across four real-world datasets, the primary validation analyses are Mozilla Perfherder’s autoregressive persistence on a real production-performance signal (R2=0.17 to 0.92) and SQuaD’s CVE-count (R2=0.23) and static-analysis-enriched defect-fix (R2=0.50) results; the confidence intervals are wide throughout, and most remaining comparisons include zero. The central limitation is that no available dataset joins eight-phase SDLC telemetry to production uptime, so each real dataset tests only a narrower proxy outcome (build duration, browser-performance regression, CI workflow duration, defect and CVE counts). PRESTO is therefore a framework and benchmark whose central proposition remains untested.