Helene
Xie

Systematic Trading

From research to live trading

I built a multi-strategy trading system that runs 8 strategies across equities, futures, and intraday timeframes. Each strategy goes through backtesting and factor attribution before it gets to run live.

I love the cycle of writing a strategy, backtesting, watching it run live, and learning from the market. The wins are fun, the losses are better teachers.

Adaptive Health Algorithms

Real-time decisions for wellbeing

At SomaBeats, I design the algorithm behind CalmingBeats — adaptive haptics on the Apple Watch. When HRV and contextual signals suggest a user is stressed, the system delivers a silent haptic pulse to help them regain calm in seconds. I built a two-head contextual bandit that learns each person's patterns over a 60-day trial, personalizing when and how to intervene.

I've been fighting my own anxiety and depression for years, so I've always wanted to use what I can build to help others going through similar things.

Portfolio Optimization

Multi-strategy allocation and risk

At China Merchants Futures, I built config-driven signal pipelines for basis and term-structure research with batch processing and multi-process daily updates. I also designed a hierarchical multi-strategy portfolio framework across CTA, market-neutral, enhanced-index, and hedging strategies.

Macro Risk Modeling

Factor-mimicking and volatility forecasting

At ICBC Credit Suisse Asset Management, I constructed macro risk factors through factor-mimicking methods across growth, inflation, rates, credit, FX, and liquidity, then used them to adjust a multi-asset risk-parity portfolio. I also built GARCH-based volatility forecasting models for dynamic risk budgeting.

In a cross-sectional stock selection project, I applied IC filtering and Ridge regression for factor combination in CSI 1000.

Audio Classification

Kaggle competition, top 10%

Classified bird species from DCT-compressed spectrograms using a 5-fold EfficientNet-B3 ensemble with MixUp, SpecAugment, and test-time augmentation. The model reached 0.54 test accuracy across 9 imbalanced classes, a 68% increase over the SVM baseline.