Paper: arXiv 2601.07131

Authors: Sungwoo Kang

Abstract

The application of machine learning to financial prediction has accelerated dramatically, yet the conditions under which complex models outperform simple alternatives remain poorly understood. This paper investigates whether advanced signal processing and deep learning techniques can extract predictive value from investor order flows beyond what simple feature engineering achieves. Using a comprehensive dataset of 2.79 million observations spanning 2,439 Korean equities from 2020–2024, we apply three methodologies: \textit{“Independent Component Analysis”} (ICA) to recover latent market drivers, \textit{“Wavelet Coherence”} analysis to characterize multi-scale correlation structure, and \textit{“Long Short-Term Memory”} (LSTM) networks with attention mechanisms for non-linear prediction. Our results reveal a striking finding: a parsimonious linear model using market capitalization-normalized flows (``Matched Filter’’ preprocessing) achieves a Sharpe ratio of 1.30 and cumulative return of 272.6%, while the full ICA-Wavelet-LSTM pipeline generates a Sharpe ratio of only 0.07 with a cumulative return of \(-5.1%\). The raw LSTM model collapsed to predicting the unconditional mean, achieving a hit rate of 47.5% – worse than random. We conclude that in low signal-to-noise financial environments, domain-specific feature engineering yields substantially higher marginal returns than algorithmic complexity. These findings establish important boundary conditions for the application of deep learning to financial prediction.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 9.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematical techniques (ICA, wavelet analysis, LSTM architectures) and includes a detailed empirical analysis with a large, multi-year dataset from the Korean stock market, rigorous backtesting metrics, and performance comparisons against simple baselines.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Does complex ML outperform<br>simple feature engineering<br>in investor flow prediction?"] --> B{"Data Input<br>2.79M Observations<br>2,439 Korean Equities<br>2020-2024"}
  
  B --> C["Method 1: ICA<br>Extract latent market drivers"]
  B --> D["Method 2: Wavelet Coherence<br>Multi-scale correlation analysis"]
  B --> E["Method 3: LSTM + Attention<br>Non-linear prediction model"]
  
  C & D & E --> F["Full Pipeline<br>ICA + Wavelet + LSTM"]
  
  G["Simple Model<br>Market Cap-Normalized Flows<br>Matched Filter Feature Engineering"] --> H
  
  F --> H["Performance Comparison"]
  G --> H
  
  H --> I["Key Findings:<br>Simple Model: Sharpe 1.30, Return +272.6%<br>Complex Pipeline: Sharpe 0.07, Return -5.1%<br>LSTM alone: Hit Rate 47.5% (worse than random)"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes that simple market‑cap‑normalized flow (Matched Filter) linear models outperform advanced ICA‑Wavelet‑LSTM pipelines for investor flow prediction.

Source sentence
We bridge these literatures by testing whether advanced signal processing (ICA, Wavelets) and deep learning (LSTM) can extract additional predictive value from investor flows beyond simple normalization.

Assumes 5

  • Transaction costs are 10 basis points per round‑trip trade. execution
    Source sentence
    Sharpe Ratio (annualized, net of 10bp transaction costs)
  • Trading strategies are evaluated on daily data with no slippage beyond the 10bp cost. execution
    Source sentence
    Net Sharpe accounts for 10bp round‑trip transaction costs.
  • Investor flow data can be normalized by market capitalization to create a scale‑invariant measure. data
    Source sentence
    Following the signal processing concept of matched filtering, we normalize flows to create a scale‑invariant measure: Sg i,t = NetBuyg i,t / MarketCapi,t
  • The ICA mixing model assumes observed flows are linear mixtures of independent latent components plus noise. model
    Source sentence
    We hypothesize that observed investor flows are mixtures of latent market drivers: Xt = A · St + ϵt
Show 1 more
  • The LSTM is trained to predict next‑day returns using mean‑squared error loss. model
    Source sentence
    Loss function: Mean Squared Error (MSE) • Optimizer: Adam with learning rate 10−3

Data

Backtest sample: 2020 – 2024 · Korean equities · equities · daily

  • Korean equities daily trading records · Korea Exchange market-prices proprietary
    2020‑2024 · 2,439 Korean equities (KOSPI and KOSDAQ)
    Source sentence
    Our dataset comprises daily trading records for all common stocks listed on the Korea Exchange (KOSPI and KOSDAQ) from January 2, 2020 through December 27, 2024.

Costs

Modeled Costs are part of the reported headline results.

10 basis points round‑trip transaction cost applied to reported Sharpe ratios.

Source sentence
Sharpe Ratio (annualized, net of 10bp transaction costs)

Finds 5

Reported performance (Simple Momentum (Linear) with Matched Filter normalization): Sharpe 1.30 · total return 272.6% · max drawdown 19.1%

  • Simple Momentum using Matched Filter normalized flows achieves a Sharpe ratio of 1.30 and cumulative return of 272.6% over five years. performance
    Source sentence
    Simple Momentum achieves a Sharpe ratio of 1.30 after transaction costs, with cumulative return of 272.6% over 5 years (annualized ∼30%).
  • The ICA‑Macro Factor strategy yields a Sharpe ratio of only 0.07 and a negative cumulative return of –5.1%. performance
    Source sentence
    ICA Macro Factor generates a Sharpe of only 0.07—essentially zero risk‑adjusted return—with a catastrophic 48.9% maximum drawdown and negative cumulative return.
  • The raw LSTM model collapses to predicting the unconditional mean, with a hit rate of 47.5%, worse than random. predictive-accuracy
    Source sentence
    The raw LSTM model collapsed to predicting the unconditional mean, achieving a hit rate of 47.5% -- worse than random.
  • Complexity destroys information: the full ICA‑Wavelet‑LSTM pipeline produces a Sharpe of 0.07 versus 1.30 for the simple linear model. negative
    Source sentence
    Our findings are unambiguous: complexity fails.
  • Normalization creates a sufficient statistic for return prediction, making additional processing detrimental. theoretical
    Source sentence
    The normalized flow may be a sufficient statistic for return prediction—containing all relevant information such that additional processing can only lose rather than gain information.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 43 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.