Métodos quantitativos para análise de criptoativos: da validação computacional à detecção de regimes em mercados emergentes

Abstract

The rise of the GameFi ecosystem has introduced market dynamics that challenge traditional financial analyses applied to crypto-assets. This dissertation investigates predictability and interdependencies between traditional cryptocurrencies and gaming tokens through an integrated production engineering analytical approach, spanning August 2013 to January 2025. The first study evaluated five machine learning algorithms forecasting Axie Infinity Shards across 1–10 day horizons. Extreme Learning Machine achieved 3.82% error for one day, deteriorating to 7.49% by day four, establishing practical limits for reliable predictability. The second study compared WEKA and R platforms using Bitcoin across 25 algorithms. M5Rules in WEKA obtained 2.25% error, while R yielded 3.79%. Severe inconsistencies in the Classify module indicated need for cross-platform validation. The third study developed a multidimensional analytical protocol for fourteen crypto-assets through parallel tracks of PCA, clustering, Hidden Markov Models, shock decomposition, Granger causality, and GARCH. Clustering identified k = 6 as optimal, revealing structural decoupling between traditional infrastructure and gaming tokens. PCA demonstrated extreme systematic risk concentration with 67.89% variance in the first component. HMM identified five regimes, with stress regime exhibiting 13.94% daily volatility. GARCH documented volatility persistence of 0.938 with 20.8-day half-life. Regularized linear models outperformed ensembles, justified by predominantly linear causal structure in 80.8% of relationships. The integrated analytical approach demonstrates that predictability limitations reflect fundamental market properties rather than methodological inadequacies.