Abstract
Bitcoin, as one of the most popular cryptocurrencies, is recently attracting much attention from investors. Bitcoin price prediction is consequently a rising academic topic for providing valuable insights and suggestions. Existing Bitcoin prediction works mostly rely on trivial feature engineering, manually designing features from multiple areas including Bitcoin blockchain information, finance, and social media sentiments. This feature engineering not only requires significant human effort, but the effectiveness of intuitively designed features cannot be guaranteed. In this paper, we aim to mine the abundant patterns encoded in Bitcoin transactions and propose the $k$-order transaction graph to reveal patterns under different scopes. We propose a transaction graph-based feature to automatically encode these patterns. A novel prediction method is proposed to accept these features and make price predictions, taking advantage of particular patterns from different historical periods. Experimental results demonstrate that the proposed method outperforms recent state-of-the-art methods.
Introduction
Predicting Bitcoin prices is notoriously difficult, often likened to catching lightning in a bottle. While many researchers look to external signals like Twitter sentiment or gold prices, this research suggests that the secret to better predictions might be hidden within the structure of the blockchain itself. By analyzing the "topology" of transaction graphs, complex transaction patterns can be automatically encoded into $k$-order subgraphs, outperforming traditional models that rely on manual feature engineering and external market indicators.
Problem Statement: The Limits of Manual Feature Engineering
Most existing Bitcoin price prediction models rely on "trivial feature engineering," where researchers manually select variables such as the number of new wallet addresses, Google Trends data, or S&P 500 performance. This approach suffers from two major flaws:
- Human Effort & Bias: Manually designing features is labor-intensive and limited by the researcher's intuition.
- Latent Information: External factors eventually manifest as changes in how users move money on the blockchain. By focusing only on external "noise," models miss the "signal" of actual transaction behavior.
The challenge lies in transforming millions of raw, messy transactions into a format a machine learning model can understand without losing the underlying structural patterns.
Methodology: Transaction Graph Analysis
The proposed framework automatically extracts features using Transaction Graphs to capture the flow of value across the network.
1. $k$-order Transaction Subgraphs
Instead of looking at a single transaction in isolation, the method examines the "flow." A 1-order subgraph shows a transaction and its immediate inputs/outputs. A 2-order subgraph follows those outputs to the next set of transactions, capturing how "gathering" (buying) or "splitting" (selling) behaviors ripple through the network.
2. Pattern Encoding via Occurrence Matrices
To make this data digestible for a model, subgraphs are categorized based on their "shape"—specifically the number of input and output addresses $(m, n)$. An Occurrence Matrix counts how often specific patterns (e.g., a 2-input, 1-output transaction) appear within a given timeframe. This turns the "topology" of the blockchain into a numerical feature vector.
3. Temporal Integration
The model uses an ensemble of learners trained on different "history windows" ($h$). These predictions are combined using a weighted decay function, prioritizing recent patterns while retaining "memory" of longer-term trends.
Datasets and Evaluation
The study utilized blockchain data from the Google BigQuery crypto_bitcoin dataset and daily closing prices from Coindesk. Models were tested on two intervals:
- Interval 1: Aug 2013 – July 2016 (1,065 days).
- Interval 2: April 2013 – April 2017 (1,461 days).
Key Results
Performance was measured using Mean Absolute Percentage Error (MAPE):
- Superior Accuracy: The integrated model achieved a MAPE of 1.69% on Interval 1, significantly beating the state-of-the-art SVM baseline of 1.91%.
- Stability: Framework performance stabilizes with a history window length of $h \ge 2$ days, indicating efficient computational requirements.
- Directional Accuracy: The model effectively captures market trends (rise or fall), providing actionable intelligence beyond simple price points.
Discussion and Insights
The core finding is that the blockchain acts as a mirror; external economic and social factors are eventually reflected in transaction structures. Furthermore, the structure of transactions (input-to-output ratios) provides deeper insight than simple volume metrics, revealing whether users are accumulating or distributing assets.
Limitations and Future Work
While promising, the method faces challenges in computational intensity when scaling history windows. Additionally, the current model caps patterns at 20 inputs/outputs to maintain matrix manageability, potentially overlooking high-impact "whale" transactions. Future research could explore real-time adaptation for high-frequency trading environments.
Summary updated