I build neural trading systems and test them on live markets.
5 papers · 2 models in production · Everything is empirical.
A multi-scale selective state space model combining Variable Selection Networks, Mamba SSM encoders, and temporal attention pooling. 2.0M parameters with O(T) complexity — 6x lighter than equivalent Transformer architectures. Drop-in replacement with identical forward signatures.
Cross-index dynamics between DJIA, S&P 500, and NAS100. Literature review complete — identifying unstudied research gaps in price-weighted divergence signals and trivariate cointegration.
Experimental results validated on live market data with walk-forward testing.
A literature review and research framework for predicting US equity index movements using cross-index dynamics. We identify several unstudied research gaps including price-weighted vs cap-weighted divergence signals and trivariate cointegration regime models. Empirical phases are in progress.
Read paper →Using the Caldara-Iacoviello Geopolitical Risk Index matched to XAUUSD M1 data (2018-2026 overlap, ~94 monthly observations), we test whether GPR predicts gold returns. GPR level regimes condition return distributions (higher volatility and positive skew in high-GPR months), but directional predictive power is weak. The signal operates at monthly frequency, too slow for intraday trading but potentially useful as a regime filter.
Read paper →We show that directional disagreement between XAUUSD and XAGUSD over a 20-bar window is the strongest single predictor of scalping signal quality, with Spearman rho between -0.23 and -0.29 (p approximately 0). Lower disagreement implies stronger co-movement and higher reversal reliability. We design a four-tier lot scaling system based on this metric, with the top tier (disagreement <= 8 plus XAG bar reversal) receiving 1.5x allocation.
Read paper →We test for linear lead-lag relationships across major asset pairs over 5.5 years of minute-level data. For gold (XAUUSD), no robust lead-lag signal exists from DXY, silver, or equity indices at any horizon. For equities, only MSFT-to-NAS100 and GS-to-US30 at the 5-minute horizon survive out-of-sample validation. The results challenge common assumptions about cross-asset predictability in systematic trading.
Read paper →Neural network architectures designed for financial time series.
We present GoldSSM, a selective state space model for intraday gold price direction forecasting. The architecture combines a Variable Selection Network, a stack of Mamba blocks with selective scan, and temporal attention pooling. At 2.0M parameters, GoldSSM serves as a drop-in replacement for Transformer-based models with identical forward signatures, while offering linear-time sequence processing and improved handling of long-range dependencies in financial time series.
Read paper →Interested in the research or exploring opportunities? Connect on LinkedIn →