ZCZifan Chen

Zifan Chen

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zifan Chen has a total of 7 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Machine Learning-Enhanced Financial Volatility Forecasting: Uncertainty-Gated Residual Boosting Across Prophet-Style, Gradient-Boosting, and Classical Time-Series Models

    by Zifan Chen

    Financial volatility forecasting is a setting in which flexible machine-learning models can exploit nonlinear interactions, yet unconstrained corrections can become unreliable when market dynamics move away from the conditions represented in the training sample. This paper develops Uncertaintyand Disagreement-Gated Residual Boosting (UDGRB), a concrete hybrid forecasting algorithm rather than a model-agnostic framework. The method first constructs a dynamic structural anchor from ARIMA, simple exponential smoothing, a heterogeneous autoregressive (HAR) model, and a transparent Prophet-style additive trend model with piecewise-linear changepoints and Fourier seasonality. A LightGBM residual learner then predicts the remaining log-volatility error. Two quantile LightGBM models estimate residual-correction uncertainty, while cross-model dispersion measures structural disagreement. A multiplicative gate shrinks the machine-learning correction according to these two signals; the gate penalties are selected only on a chronologically separate validation period and are allowed to be exactly zero, yielding data-driven sparsification or fallback when a gating signal has no incremental value. The empirical study uses 7,983 real Microsoft daily OHLCV observations from 13 March 1986 to 10 November 2017, distributed with the public pmdarima dataset and traceable to the Kaggle US stocks/ETFs price-volume collection. We form 1,598 leakage-controlled weekly forecasting observations after lag construction and evaluate 567 one-week-ahead forecasts from 2007-2017. The primary target is annualized close-to-close realized volatility; a Parkinson high-low range estimator supplies an alternative volatility proxy. Baselines include naive persistence, ARIMA, exponential smoothing, HAR, the Prophet-style additive model, standalone LightGBM, and dynamic forecast combination. On close-to-close volatility, UDGRBachieves the lowest MAE (8.292) and MASE (0.610), improving MAE by 3.09% over standalone LightGBM; the Diebold-Mariano test for absolute-error loss gives 𝑝 = 0.0427. On the Parkinson target, the validation criterion selects the no-shrinkage corner, so the method automatically reduces to the residual hybrid; it again obtains the lowest MAE (7.177). HAR produces the best variance-scale QLIKE on both targets, showing that the ranking is loss-dependent rather than universally dominated by one model. Importantly, on the primary target the uncertainty gate reduces QLIKE relative to the ungated residual hybrid by 0.0241 on average, with an eight-week moving-block bootstrap 95% interval of [-0.0640,-0.0003]. The experiments therefore identify a specific role for uncertainty-aware residual shrinkage: it improves point accuracy and protects residual boosting when validation evidence supports intervention, while preserving a statistically stronger structural solution when it does not.

  • Top Cited Papers
    Open Access

    Uncertainty-Aware Financial Forecasting: Leveraging Conformal Prediction for Risk-Adjusted Models

    by Zifan Chen, Mengdie Wang, Zhuoqi Zeng, Wenjie Ping

    Financial forecasting presents a critical challenge in risk management, where traditional point estimation approaches fail to capture the inherent uncertainty and volatility of financial markets. While deep learning models have demonstrated remarkable predictive capabilities, they typically provide deterministic predictions without quantifying prediction confidence or risk boundaries. This paper introduces a practical uncertainty-aware forecasting framework that integrates Conformal Prediction (CP) with deep recurrent neural networks to generate statistically valid prediction intervals alongside point estimates. Our approach leverages Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures as base learners, enhanced with adaptive conformal inference mechanisms that provide approximate marginal coverage under realistic market dynamics. We propose three key design contributions: (1) a volatility-adaptive nonconformity measure that adjusts prediction intervals based on market regime detection, (2) a temporal split conformal framework specifically designed for financial time series with autocorrelation structures, and (3) an extreme event awareness mechanism that dynamically widens intervals during high-uncertainty periods. Comprehensive experiments on multiple stock market datasets spanning multiple years demonstrate that our framework achieves superior calibration (95.2% empirical coverage at 95% target level) while maintaining competitive predictive accuracy (RMSE reduction of 8.3% compared to baseline models). Notably, our method exhibits enhanced early warning capabilities for extreme market events, achieving 87.6% detection rate for significant price movements with controlled false positive rates. We additionally evaluate the framework on cryptocurrency data and with multiple modern backbone families across the same 10-stock subset, providing empirical evidence that VAC-FF is compatible with diverse base predictor architectures and asset classes. The proposed framework offers financial practitioners a transparent, statistically rigorous tool for risk-adjusted decision-making, bridging the gap between predictive accuracy and uncertainty quantification in high-stakes financial applications.


  • Top Cited Papers
    Open Access

    LLM-Driven Adaptive Cloud Resource Scheduling: Bridging Reasoning Intelligence With Optimization Guarantees

    by Guanyu Ding, Shiyu Yang, Han Lin, Zifan Chen, Jie Si Yang

    Cloud resource scheduling presents a fundamental challenge in modern distributed computing, where heterogeneous workloads, complex task dependencies, and multi-objective optimization requirements exceed the capabilities of traditional rule-based and small-scale machine learning approaches. Existing schedulers struggle to dynamically adapt to evolving workload patterns while simultaneously satisfying Service Level Agreement (SLA) constraints, resource efficiency targets, and fairness policies. This paper introduces LLMSched, a novel Large Language Model (LLM)-driven adaptive scheduling framework that synergistically combines the contextual reasoning capabilities of foundation models with the execution guarantees of classical optimization techniques. Our approach transforms heterogeneous cluster states-including task dependency graphs, real-time resource utilization metrics, and SLA specifications-into a unified structured-textual representation that leverages LLMs' few-shot learning and causal reasoning abilities to generate intelligent scheduling candidates. These candidates are subsequently refined through a lightweight Integer Linear Programming (ILP) module that ensures feasibility and optimality under resource constraints. We evaluate LLMSched on Google's production cluster trace dataset, demonstrating significant improvements over state-of-the-art baselines: 23.7% reduction in average job completion time, 18.4% improvement in resource utilization efficiency, and 31.2% decrease in SLA violations across diverse workload scenarios. Extensive ablation studies validate the contributions of each architectural component, while robustness analysis under workload perturbations confirms the framework's practical viability. Our work establishes a new paradigm for intelligent resource management that bridges the gap between neural reasoning and algorithmic precision, opening avenues for LLM applications in systems optimization domains.