Rahul S. P.

Quantitative research on financial time series: what is real, what is a mirage, and how to tell the difference.

8 papers · Everything is empirical.

LinkedInAbout
Live nowInteractive globe

World Money-Flow Map

Watch where capital is rotating right now: gold, crypto, bonds, the dollar and equity indices across the US, Europe and Asia. Rebuilt every minute from measured fund flows, order-flow and price rotation.

Open the live globe
FeaturedArchitecture & Models

Transformer Models vs SSMs for Financial Time Series

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.

Mamba SSMVariable SelectionO(T) Complexity2.0M ParamsRead paper →
In ProgressEmpirical Studies

US Index Prediction: A Multi-Index Framework

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.

Phase 1: Literature ReviewUS30US500NAS100Read more →

Empirical Studies

7

Experimental results validated on live market data with walk-forward testing.

How Much, Not Which Way: What 106 Public Features Actually Predict About the Dow

We set out to forecast the direction of the Dow (US30) and, after trying every model, target, and horizon, kept landing on the same flat result: a profit factor of 1.0. Instead of re-tuning, we set the trade engine aside and ran an engine-independent battery to separate the five distinct causes of a flat equity curve. Across 438,619 out-of-sample bars and 16 walk-forward folds, our features (all public, mostly price-derived) predict 60-minute realised volatility with a rank correlation of 0.70 (36 sigma above a shuffled-label null) and predict direction no better than a coin (IC 0.009, accuracy 50.8%, 0 of 106 features surviving false-discovery control). The result holds from 60 minutes to 20 days, survives a linear and a neural alternative, and is confirmed by a label-shuffle retrain null. The conclusion was not a weak model but a wrong target: size by predicted volatility, and stop forecasting direction from price.

Read paper →

Three Textbook Edges Meet the Spread: A Mean-Reversion Replication on Index CFDs

We replicate three of the most-cited OHLCV-only mean-reversion edges on the instruments people actually trade (US30, US500 and NASDAQ CFDs), with a realistic spread on every entry and exit and a frozen-parameter walk-forward. Overnight gap-fade does not replicate: its famous ~90% fill rate is an artefact of tiny gaps, and gaps large enough to fade fill under half the time (only 16% for large NASDAQ gaps). The IBS effect's reported ~70% win rate collapses to ~50%, a coin flip. Connors RSI(2) is the one partial survivor, with real 57-67% win rates, but it trades too rarely to beat a trending index. After costs and out-of-sample, none of the three beats buy-and-hold. The point is not that mean reversion is fake, but that venue, costs, and out-of-sample decay separate a backtest in a paper from a position in a live account.

Read paper →

Geopolitical Risk and Gold: An Empirical Study

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 →

XAG Directional Disagreement as a Cross-Asset Lot Scaling Signal

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 →

A Real Edge that Retail Can't Trade: A Cross-Sectional Factor Study on 1,021 US Stocks

After a separate study found the direction of a single index to be a coin flip, we asked whether the cross-section of stocks holds the information the level does not. On 1,021 US names from 2005 to 2026, with point-in-time SEC EDGAR fundamentals and an embargoed, expanding, cost-aware walk-forward, the cross-sectional signal is genuinely real: twelve-month momentum at t = 2.76, five-day reversal at t = 2.62, and the value premium showing its two-decade drought (earnings yield t = -4.14). Yet the combined long/short book earns a net Sharpe of just 0.38, below the 0.46 deflated-Sharpe hurdle that corrects for multiple testing, and its sector-neutral refinement is worse at 0.20. The naive momentum-only sleeve lost 71% in a single month in 2009 and 96% peak-to-trough. The signal is real and, for a retail account that cannot short hundreds of names against an institutional borrow desk on survivorship-free data, untradeable. The distance between a significant t-statistic and a dollar in a real account is the subject of the paper.

Read paper →

Architecture & Models

1

Neural network architectures designed for financial time series.

Transformer Models vs SSMs 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 →