Selected work

05 · Time-series machine learning

Bitcoin Direction Model

An exploratory three-day Bitcoin-direction study that prioritizes chronological validation and uncertainty over headline accuracy.

observations4,062
features38
test accuracy53.93%

chronological split · TimeSeriesSplit · bootstrap interval

Role

Independent quantitative research - data preparation, feature engineering, model comparison, and reporting.

Context

The study combines Yahoo Finance history and the CoinMetrics Community API, then filters near-flat forward-return observations before model evaluation.

Tools

Python · XGBoost · scikit-learn · TimeSeriesSplit · NumPy · pandas · Yahoo Finance

Executive summary

A quantitative research workflow built from market and on-chain data to test whether engineered signals improve directional classification on unseen data.


Problem

Financial time-series prediction is prone to look-ahead bias. The project tests several algorithms while preserving chronological splits and making performance uncertainty visible.

Approach

  1. 01Merge daily price history and on-chain MVRV data, then construct momentum, trend, volatility, oscillator, lag, and valuation features.
  2. 02Use a three-day forward-return target and remove near-flat movements under a stated ±1% filter.
  3. 03Train linear, logistic, Lasso, random-forest, and XGBoost models using chronological splits and TimeSeriesSplit validation.
  4. 04Use bootstrap confidence intervals to contextualize directional test accuracy rather than treating one score as conclusive.
02

Data & methodology

Data sources

  • Yahoo Finance daily BTC data
  • CoinMetrics Community API MVRV series

Methods

  • 4,062 usable daily observations before target filtering
  • 38 engineered market and on-chain features
  • Chronological 80/20 split with TimeSeriesSplit
  • 1,000-resample bootstrap confidence intervals
Workflow
InputCollectionAnalysisOutput

Results are reported as documented in the project report and are presented with their study constraints.

Outputs & findings

Key outputs

  • Executive research report
  • Model comparison
  • Feature-engineering specification
  • Validation and confidence-interval analysis

What the work demonstrates

  • 3,178 observations remained after the project’s stated near-flat-return filter.
  • XGBoost reached 53.93% directional test accuracy; its reported 95% bootstrap interval was [50.31%, 57.70%].

Limitations

  • Exploratory research, not a live trading strategy or investment recommendation.
  • Directional accuracy alone does not establish tradability, robustness across regimes, or net performance after costs.
  • The filter and data period materially shape the evaluation set.