RECOR Stock Prediction Framework — Full Breakdown & Transcript

【论文分享】库普曼+扩散残差择时

0h 12m video Published Jul 11, 2026 Transcribed Sep 14, 2026 量化Quantgirl

3 views Recent velocity 0.0 views/hour View full performance history →

Advanced 6 min read For: Quantitative researchers, machine learning practitioners, and finance professionals with a background in time series modeling and deep learning.

AI Trust Score 75/100

⚠️ Average / Some Fluff

"The title promises a novel method and delivers exactly that—a detailed technical breakdown of the RECOR framework with empirical results."

AI Summary

This video presents a research paper on a novel stock prediction framework called RECOR (Residual-Enhanced Adaptive Koopman Autoencoder). The framework combines Koopman operator theory with conditional diffusion models to address the non-stationary and nonlinear nature of financial time series, achieving improved performance on CSI 300 and S&P 500 indices.

Chapters

1 Introduction and Motivation 00:00 2 Core Contributions and Problem Setup 01:03 3 Architecture Details 02:36 4 Diffusion Residual Correction 06:34 5 Experiments and Results 08:17 6 Conclusion and Future Work 10:30

[00:00]

Introduction to RECOR

The video introduces a residual-enhanced adaptive Koopman autoencoder (RECOR) for stock prediction, highlighting its application to CSI 300 and S&P 500 indices.

[00:37]

Core Problem

The key challenge is modeling changing market states from a dynamical systems perspective while capturing shocks and nonlinear residual effects.

[01:03]

Koopman Theory

Koopman theory lifts complex nonlinear systems into a space where evolution becomes linear, but finite approximations leave residuals containing higher-order nonlinear structures.

[01:43]

Four Core Contributions

The paper's contributions include adaptive latent dynamics, residual enhancement with conditional diffusion, dual-channel representation, and empirical gains in annualized return and Sharpe ratio.

[02:36]

Problem Formulation

Stock prediction is framed as a sequence-to-sequence task: given historical features and returns, predict future returns, evaluated using rank correlation (RANKC).

[03:09]

Architecture Overview

The framework includes a dual-path LSTM encoder, gating module, adaptive Koopman selector, dynamic residual corrector, and decoder.

[04:04]

Latent Embedding

LSTM encoders produce latent embeddings stacked over time to form snapshot matrices, enabling linear evolution via Koopman operators.

[05:11]

Adaptive Koopman Selector

A codebook of multiple Koopman operators is used, with a selector network choosing the appropriate operator based on current latent state, using Gumbel-Softmax for differentiable selection.

[06:34]

Dynamic Residual Corrector

A conditional diffusion model generates residuals conditioned on the latent state, capturing higher-order nonlinearities and shocks missed by Koopman operators.

[08:17]

Experimental Setup

Datasets: CSI 300 and S&P 500 from 2010-2020, using Alpha 158 features from QLIB, with a 10-day lookback window.

[09:11]

Performance Results

RECOR achieves top performance on RANKC and RANKCS across both markets, significantly outperforming the MASTER baseline.

[09:53]

Ablation Study

Removing any of the three key modules (adaptive selector, diffusion corrector, gating) degrades performance, confirming their importance.

[10:30]

Investment Simulation

Top-K strategy: hold top 30 stocks based on predicted scores, rebalanced daily; RECOR shows robust cumulative returns.

[10:56]

Conclusion and Future Work

Future directions include richer market state representations, incorporating macro variables, and extending to high-frequency trading.

RECOR provides a principled and effective dynamical systems approach to stock prediction, combining Koopman theory with diffusion-based residual correction, and demonstrates significant improvements over existing methods.

Mentioned in this Video

QLIB

tool

💡 Key Takeaways

🔧

Koopman Theory Application

Introduces a principled dynamical systems approach to stock prediction, moving beyond traditional sequence models.