Market represented as a time-indexed multi-layer graph G_t with separate layers for correlation, sector/factor exposure, and sentiment co-movement; current experiments use only the correlation layer to isolate topology effects .
Correlation edges: Spearman correlations computed on a short rolling window (implementation uses W=7 days), edges included if |ρ_ij| ≥ τ with τ=0.5 producing sparse, responsive networks; node features are light (F=7) to emphasize structure over complex features .
Temporal GNN prototype: sequence of snapshot GCN embeddings (k=5 daily graphs) aggregated by a single-layer GRU, logistic output; prototype purposely simplified (no multi-layer fusion, short sequences) to provide a first temporal baseline .
Quantitative performance summary (key metrics)
Aggregate observation: no model uniformly dominates; temporal GNN improves AUROC on Dot‑com relative to snapshot GNN, but correlation-only GNNs fail on COVID‑19 due to exogenous dynamics not reflected by pre-crash correlations .
Reproduced core performance table (selected rows from report) :
Crisis
Model
AUROC
Precision
Recall
Accuracy
Dot-com
Logistic Regression
0.463
0.592
0.438
0.470
Dot-com
Random Forest
0.291
0.601
0.948
0.581
Dot-com
GNN
0.412
0.613
1.000
0.613
GFC
Logistic Regression
0.738
0.824
0.977
0.808
GFC
Random Forest
0.529
0.827
1.000
0.827
GFC
GNN
0.248
0.823
1.000
0.823
COVID-19
Logistic Regression
—
1.000
0.823
0.823
COVID-19
Random Forest
—
1.000
0.967
0.967
COVID-19
GNN
0.000
1.000
1.000
1.000
Interpretation: correlation-only GNNs can capture structural buildup (Dot‑com, partially GFC) but yield degenerate / inverted signals when crises are exogenous (COVID‑19), motivating multi-layer (sector/sentiment) inputs .
Lead-time & interpretability findings
Lead-time histogram indicates many warnings precede drawdowns by multiple weeks for Dot‑com and GFC, but distributions vary by crisis, underscoring regime dependence of early-warning lead-times .
Feature importance (Random Forest) highlights volatility, average correlation, and drawdown as top predictors; GNN attention maps focus on liquid sector ETFs, indicating emphasis on meso/macro structure rather than idiosyncratic signals .
Rolling window for correlation edges: W = 7 days, Spearman correlation, threshold τ = 0.5 to form edges .
Node features: 7 features (daily returns; 20/60-day volatility; 20/60-day drawdowns; 10/30-day momentum) and simple macro overlays used as graph-level context .
Temporal aggregation: sequence length k = 5; GCN encoder produces Z_t per graph; GRU aggregates (single layer) → logistic output; loss = BCE (chronological 80/20 split, Adam lr=1e-3) .
Evaluation metrics: AUROC, AUPRC, precision, recall, accuracy and empirical lead-time distributions; class imbalance handled by reporting precision–recall rather than calibration in this stage .
思维导图
深度阅读
元数据与概览(引言与报告概览)
报告标题与作者:文档标题为 “Systemic Risk Radar: A Multi-Layer Graph Framework for Early Market Crash Prediction”,作者为 Sandeep Neela(Independent Researcher),联系邮箱在文首注明。
推理与假设:核心假设为“结构性不稳定性(例如相关性上升、波动率聚集、行业同步)在价格大幅下跌之前可检测到”,并据此构建多层图来分离不同传播渠道(相关性、行业暴露、情绪)。该观念与文献(Battiston, Haldane & May 等)关于网络连通性和放大效应的理论一致,因此作者以网络拓扑变化作为系统性信号的代理。