← Research Lab / Applied
FIG. A3BE // FANTASMA RESEARCH NOTE

Signal Integrity in High-Frequency Financial Inference

Oleh Yuki Tanaka
12 min baca
2025-06-08

Signal Integrity in High-Frequency Financial Inference

Financial ML has a problem that most domains don't: the model's own predictions affect the data distribution it's trained on. A model that correctly identifies an alpha signal will eventually cause that signal to be arbitraged away. If you don't account for this, your validation metrics are systematically optimistic.

We spent 18 months building a production inference system for a tier-1 quantitative trading desk. Here's what we learned.

The Feedback Problem

Standard ML assumes i.i.d. data. Financial time series is anti-i.i.d. — serial correlation, regime changes, and reflexivity (Soros's term) mean that the distribution at inference time is causally downstream of the model's previous predictions.

This manifests as overfitting to stale alpha: a model trains well, goes live, makes money, and within 60–120 days has arbitraged its own edge away.

Our Approach

We built a feedback-aware training loop that:

Simulates the market impact of the model's predicted trades during training
Re-computes the training labels under the impact-adjusted prices
Iterates until convergence (typically 3–5 rounds)

This isn't a novel idea — it's related to equilibrium modeling in game theory. The engineering novelty is making it tractable at the data scales involved (2TB of tick data, 8M simulated trades per round).

Results

Live alpha half-life improved from 73 days to 194 days. Sharpe ratio degradation over 6 months: -12% (vs. -58% for baseline).

Publikasi Riset Terkait

Architecture 11 min

Attention Is Not Enough: Toward Structured World Models at the Edge

Transformer attention scales quadratically with sequence length. For always-on edge inference, this is a thermodynamic impossibility. We propose a hybrid SSM-attention architecture that achieves 94% of full-attention quality at 8% of the compute.

HCI 8 min

Calibration-Free Gaze Tracking at 60fps on Commodity Hardware

Classical gaze tracking requires a 90-second calibration ritual that most users abandon. We trained a universal gaze model on 4.2M synthetic eye-renders that generalizes to unseen users with 0.4° accuracy — no calibration, no personal data.