Paper: arXiv 2308.08558

Authors: Minsuk Kim, Byungchul Kim, Junyeong Yong, Jeongwoo Park, Gyeongmin Kim

Abstract

Financial time series have historically been assumed to be a martingale process under the Random Walk hypothesis. Instead of making investment decisions using the raw prices alone, various multimodal pattern matching algorithms have been developed to help detect subtly hidden repeatable patterns within the financial market. Many of the chart-based pattern matching tools only retrieve similar past chart (PC) patterns given the current chart (CC) pattern, and leaves the entire interpretive and predictive analysis, thus ultimately the final investment decision, to the investors. In this paper, we propose an approach of ranking similar PC movements given the CC information and show that exploiting this as additional features improves the directional prediction capacity of our model. We apply our ranking and directional prediction modeling methodologies on Bitcoin due to its highly volatile prices that make it challenging to predict its future movements.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper uses relatively standard machine learning techniques (XGBoost) and technical analysis indicators without advanced mathematical derivations, but features a complete implementation with specific backtest-ready steps, dataset details, and performance metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve Bitcoin price direction prediction using multimodal pattern matching"] --> B["Data Collection & Preparation<br>Historical Bitcoin OHLCV data"]
  B --> C["Key Methodology: PC Ranking<br>Rank similar past chart patterns to current chart"]
  C --> D["Feature Engineering<br>Generate directional features from ranked PCs"]
  D --> E["Computational Process<br>Input multimodal features into predictive model"]
  E --> F["Key Outcome: Directional Prediction<br>Enhanced model predicts future market direction"]
  F --> G["Conclusion<br>Exploiting similar past patterns improves predictive capacity"]

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 four ranking methodologies for chart pattern matching and a BTC directional forecasting model that uses voting information from past matched patterns.

Source sentence
We propose four different ranking methodologies for chart pattern matching and rank similar PC segments based on proposed metrics.

Assumes 2

  • No take profit and a stop loss of 0.75% are assumed for backtesting. execution
    Source sentence
    We assume no take profit and a stop loss of 0.75%.
  • A commission rate of 0.04% (maker fee) is applied to trades. execution
    Source sentence
    We use commission rate of 0.04%, equivalent to the maker fee when trading BTC/USDT perpetual in Binance.

Data

  • 4‑hourly BTC/USDT data · Binance market-prices public
    2017-08-23 16:00:00 to 2023-01-15 20:00: · Bitcoin (BTC) against USDT
    Source sentence
    We collected 4 hourly BTC/USDT data from Binance 3, one of the largest crypto exchanges, and we ended up with 11,812 data points ranging from 2017-08-23 16:00:00 to 2023-01-15 20:00:00.

Recognised sources: Binance

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Partial

Paper text extraction: Modeled - Commission rate of 0.04% per trade is included in the backtest.

Source sentence (the authors' words)
We use commission rate of 0.04%, equivalent to the maker fee when trading BTC/USDT perpetual in Binance.

Finds 3

  • All IR‑based feature engineering (FE) improves over the baseline with the multimodal strategy performing best. performance
    Source sentence
    On average all of our IR-based FE improves over the baseline of no FE, with the multimodal strategy performing the best.
  • Using the top 5 most similar multimodal embeddings yields the highest F1 score of 0.628. predictive-accuracy
    Source sentence
    Using the top 5 most similar multimodal embeddings has the best F1 score of 0.628
  • The multimodal model outperforms a buy‑and‑hold strategy over the test period. performance
    Source sentence
    our model outperforms buy and hold for the duration of the test set.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 9 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.