Relevance
8/10
Directly addresses cross-company news propagation, a key alpha source in event-driven and statistical arbitrage strategies. The contagion intensity vector can serve as an additional feature for downstream forecasters or risk systems. The 1.70x precision lift over random and 3.36x over sector baseline at extreme return thresholds demonstrate actionable signal. The intensity-weighted portfolio ordering produces positive returns with 57.1% daily win rate. The architecture's microsecond ingestion and millisecond inference are compatible with intraday trading workflows. However, evaluation is limited to daily resolution and a small corpus, and the system is complementary to (not a replacement for) per-ticker forecasting models.
Implementation Complexity
8/10
High complexity due to heterogeneous Rust/Python architecture with inter-process communication, continuous-time Neural Hawkes Process with per-node c-LSTM states, bilinear latent projection, adaptive edge pruning with formal guarantees, maximum likelihood training with compensator integral approximation, and multi-source graph construction (co-mention, semantic similarity, return correlation). Requires expertise in systems programming (Rust, zero-copy parsing, Criterion benchmarking), deep learning (PyTorch, LSTM, attention mechanisms), stochastic processes (Hawkes processes, point process likelihood), and graph algorithms. However, the code is publicly available and runs on a single commodity CPU without GPU.
Reproducibility
5/5
All source code, trained model checkpoints, and evaluation scripts are publicly available at https://github.com/kcbir/zcsc. Fixed random seed (42), deterministic PyTorch (torch.use_deterministic_algorithms(True)), CPU-only execution for bitwise reproducibility, Criterion benchmarks for Rust, detailed hyperparameter sensitivity analysis (Appendix C), and full software stack specification (Python 3.11, PyTorch 2.2, Rust 1.77). Temporal holdout protocol clearly defined with no leakage.