KOSUNI CRYPTO
Korea's Crypto Pulse, in English
$BTC $ETH $XRP $SOL $DOGE
Market Analysis
Neutral

Bitcoin Prediction Models Struggle to Outperform Simple Baselines

Published September 25, 2026 9:05 AM · 14 views $BTC
Bitcoin Prediction Models Struggle to Outperform Simple Baselines

Bitcoin (BTC) price prediction models have expanded to include scarcity metrics, on-chain indicators, power laws, and machine learning, yet repeated real-world testing shows they struggle to outperform simple baselines that merely use today's price or zero return. CryptoSlate summarized the Bitcoin prediction model competition, emphasizing all models should first be compared against a naive baseline. Complex models may convincingly explain past prices but do not necessarily demonstrate superior short- and medium-term forecasting ability.

Carlos Baquero highlighted in his 2026 arXiv paper that it is difficult to find peer-reviewed research consistently surpassing the naive baseline across various market conditions for 1–6 month intervals. The paper argues the core issue in Bitcoin prediction research lies not in model scarcity but in evaluation methods. A naive baseline refers to the simplest possible forecast, such as today's price or zero return. In markets moving close to random walk patterns, surpassing this baseline is inherently challenging. Prediction models must first prove superior performance against this simple benchmark before claiming complexity.

The problem is that Bitcoin markets are not fixed laboratory environments; participant behavior, liquidity, regulatory landscapes, and trading structures constantly change, causing rules trained on past data to lose effectiveness in new phases. Baquero also identified issues like non-stationarity, overfitting during backtesting, and information leakage risks. His proposed validation methods include walk-forward testing, multiple holdout periods, naive baseline comparison, and incorporating transaction costs.

Explanation-based models face similar limitations. Austin Shelton concluded in a 2024 paper that stock-to-flow and Metcalfe's Law explained Bitcoin returns within sample data but lacked predictive power outside the sample. Stock-to-flow measures scarcity by dividing existing supply by new supply, while Metcalfe's Law interprets price based on network value proportional to users squared. Both models may describe long-term structures but do not reliably predict short-term price movements.

Machine learning is no exception. A 2026 Frontiers in Blockchain study using over 3 million minute-level data points and Binance spot/Perpetual futures data found that after enhancing leakage control, LightGBM models exhibited severe overfitting, and when realistic transaction fees and slippage were factored in, no strategy survived. Simple OLS models showed only marginal improvement over random walk.

The power law debate remains open. Giovanni Santostasi and Stephen Perrenod claimed in a 2026 Nonlinear Science paper that Bitcoin dollar prices followed a power law from July 2010 to February 2026, citing an index of 5.69, R-squared of 0.961, and stable residuals. However, this describes long-term price structure—not short-term directional accuracy. Market structures require separate validation of statistical accuracy and economic profitability; high frequency prediction does not guarantee actual trading success due to transaction costs, slippage, funding rates, and execution delays.

For Korean investors, understanding the distinction is critical: models offering long-term trend lines or on-chain indicators can serve as reference tools but should not be directly applied as short-term trading signals without out-of-sample validation and cost consideration. The Bitcoin model debate has shifted from 'sophisticated formulas' to 'verifiable predictions,' emphasizing that explanatory power, predictive ability, long-term trends, short-term directionality, statistical accuracy, and actual trading performance are distinct concepts.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: 비트코인 예측 모델, 단순 기준선 넘기 어렵다