Abstract

The proposed framework, Momentum AI, evaluates Naïve persistence baselines, ARIMA

models, and Long Short-Term Memory (LSTM) networks under strict chronological validation

and preprocessing steps designed to prevent look-ahead bias. In parallel, a transparent

Momentum Fear and Greed Index (MFGI) is constructed using three explicitly defined

components: price momentum, realized volatility, and aggregated daily news sentiment

derived from publicly available cryptocurrency headline data. All normalization procedures

rely on expanding windows to prevent look-ahead bias, and sentiment features are temporally

lagged to preserve causal ordering.

Empirical results demonstrate that increasing model complexity alone does not over-

come structural forecasting limitations. While nonlinear models capture short-run temporal

structure, all approaches experience substantial degradation during shifts between market con-

ditions. The inclusion of sentiment-derived features helps the framework better reflect broader

market conditions and makes market behavior easier to interpret, though improvements in

predictive accuracy remain conditional rather than universal.

By emphasizing transparency, careful chronological evaluation, and contextual modeling

that accounts for changing market conditions, this thesis clarifies the structural limits of

short-horizon prediction in digital asset markets while offering a structured framework for

sentiment-aware analysis.

Advisor

Guarnera, Heather

Department

Computer Science

Disciplines

Artificial Intelligence and Robotics | Computer Sciences | Numerical Analysis and Scientific Computing

Publication Date

2026

Degree Granted

Bachelor of Arts

Document Type

Senior Independent Study Thesis

Download

Share

COinS
 

© Copyright 2026 Selorm Y. Essey