Research map / Graph learning / Knowledge and temporal graphs
ECHO-GL: Earnings Calls-Driven Heterogeneous Graph Learning for Stock Movement Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ECHO-GL, which builds an earnings-call-driven heterogeneous dynamic graph (E-Graph) and combines heterogeneous graph attention with a learnable SDE modeling post-earnings-announcement drift to predict multi-horizon stock movements.
Problem
Stock movement prediction relies on stock relations that are either insufficient or static (price-based or wiki/industry graphs), failing to capture dynamic inter-stock relations; also needs predictions across multiple time horizons after earnings calls.
Approach
Builds a time-stamped heterogeneous graph (price, sentence, topic, entity nodes) from earnings call transcripts with time assignment and sliding-window mechanisms, aggregates with a heterogeneous graph attention network using audio-guided attention, and models post-earnings-announcement drift with a learnable neural SDE to extrapolate stock representations over arbitrary horizons.
Key results
- Averaged over horizons, ECHO-GL improves over best baselines by 2.297% F1 and 15.629% MCC (text reports across both datasets).
- On Qin's dataset, ECHO-GL reaches F1 63.5% at 1-day horizon and MCC 30.7% at 3-day, versus best baseline MCC around 16-19%.
- Trading simulation (Oct 2017 to Feb 2018) shows positive profits on all horizons 1-30 days and annual Sharpe ratio exceeding the S&P 500 benchmark.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ECHO-GL instead of the strongest baseline, Ensemble | stock movement prediction · Qin’s, 1-day | F1: 62.7 ± 1.7 -> 63.5 ± 0.3 | Helped |
| use ECHO-GL instead of the strongest baseline, Ensemble | stock movement prediction · Qin’s, 1-day | MCC: 16.0 ± 2.4 -> 20.2 ± 4.4 | Helped |
| use ECHO-GL instead of the strongest baseline, NumHTML | stock movement prediction · MAEC, 1-day | F1: 53.4 ± 1.5 -> 54.4 ± 0.7 | Helped |
| use ECHO-GL instead of the strongest baseline, HTML | stock movement prediction · MAEC, 1-day | MCC: 6.4 ± 2.9 -> 6.7 ± 0.2 | Helped |
| add entity nodes to the graph (ablation) | ECHO-GL · Qin’s and MAEC | stock movement prediction performance | Helped |
| add topic nodes to the graph (ablation) | ECHO-GL · Qin’s and MAEC | stock movement prediction performance | Helped |
| add the stock spatial relational module (HGAN) (ablation) | ECHO-GL · Qin’s and MAEC | stock movement prediction performance | Helped |
| add the post earnings call stock dynamics module (ablation) | ECHO-GL · Qin’s and MAEC | stock movement prediction performance | Helped |
| use an earnings-call-driven heterogeneous graph instead of price or wiki graphs (ablation) | ECHO-GL · Qin’s and MAEC | stock movement prediction performance | Helped |
Limitations
- Potential biases in earnings calls (gender, demographic) not addressed
- Evaluated only on two datasets with US stocks
- Trading simulation covers a short period (Oct 2017 to Feb 2018)
Experimental setup
Datasets: Qin's earnings call dataset (S&P 500) MAEC earnings call dataset (S&P 1500) Yahoo Finance dividend-adjusted closing prices
Benchmarks: Qin's dataset MAEC dataset
Metrics: F1 MCC profit ($) Sharpe ratio
Baselines: ARIMA LSTM BiLSTM Transformer Informer Autoformer Convtrans RSR VolTAGE STHAN-SR HISN MDRM
Builds on: Neural ODE HGAT-style heterogeneous attention LDA FinBERT PEAD
Closest papers by meaning
- MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction (AAAI 2024 · similarity 0.86)
- CausalStock: Deep End-to-end Causal Discovery for News-driven Multi-stock Movement Prediction (NeurIPS 2024 · similarity 0.85)
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (AAAI 2024 · similarity 0.84)
- CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph (AAAI 2024 · similarity 0.83)
- StockMixer: A Simple Yet Strong MLP-Based Architecture for Stock Price Forecasting (AAAI 2024 · similarity 0.83)
- DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock Prediction (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.