Hyperstock technical architecture revealed

overall architecture

User → CDN → Front-end (React) → API Gateway → Service layer → Data source ↓ ↓

PostgreSQL

↓

Redis(cache)

Deployment: AWS + Alibaba Cloud active-active (overseas users use AWS, domestic users use Alibaba Cloud)

AI model selection

Score Scoring model

It's not deep learning, it's a multi-factor weighting model.

Why not use deep learning?

1. Interpretability: Users ask "Why is this Score high", and I can clearly explain the weight of each dimension.
2. Data volume: Although there is a lot of option data, the effective annotation data is not enough to train complex models.
3. Iteration speed: 5 minutes to go online to adjust weights, 5 hours to retrain the model
4. Stability: Linear models do not suddenly go crazy like neural networks

type structure:

Score = w1×annualized return_score + w2×exertion probability_score +w3×margin efficiency_score + w4×volatility_score + w5×time decay_score where w1=w2=0.25, w3=w4=0.20, w5=0.10

The score of each dimension is a piecewise linear function, not a simple linear function, taking into account the nonlinear characteristics in actual transactions.

May be introduced in the future:

• Machine learning assisted parameter adjustment (automatic optimization of weights)
• User behavior feedback closed loop (which recommended users actually use it, fed back to the model)

data source

Options real-time data:

Why choose OPRA:

• Official data source for U.S. options markets
• Full market coverage (NASDAQ/NYSE/AMEX)
• Including complete fields such as Greeks, IV, trading volume, etc.

Data processing flow:

OPRA real-time streaming → data cleaning → anomaly detection → storage → API response

↓

Abnormal data alarm (such as price jump > 50%)

SPAN Margin calculation

This is Hyperstock’s unique skill. Most options instruments don't do this.

Implementation method:

1. Parse CME SPAN parameter file (updated daily)
2. Simplified calculations for individual stock options (not complete futures portfolio SPAN)
3. Consider three risk scenarios: underlying price fluctuations, IV changes, and expiration time
4. Output estimated margin

Discrepancies with mainstream brokerages:

Why there is an error:

• Different brokers add different buffers based on SPAN
• The position portfolio has a net margin effect (we only count single contracts)
• Under extreme market conditions, brokers may temporarily increase margins

Performance optimization

Goal: Results within 3 seconds

Optimization means:

Actual performance:

• P50 response time: 1.2s
• P95 response time: 2.8s
• P99 response time: 4.1s (very few complex targets)

Bottleneck: When performing multi-standard batch analysis (10 targets at the same time), the calculation time increases linearly. The next optimization step: parallel computing.

Security and Compliance

Data security:

• Transport: Site-wide HTTPS, TLS 1.3
• Storage: PostgreSQL AES-256 encryption
• Password: bcrypt hash, random salt
• API Key: JWT + expiration mechanism

Service security:

• DDoS Protection: Cloudflare
• WAF: Custom rules to filter malicious requests
• Current limiting: Press API Key to limit current and prevent brushing

Compliance:

• User transaction data is not stored (only orders placed at brokers)
• No sensitive personal information is collected
• Risk warnings cover key pages of the entire site
• Disclaimer Users must confirm when registering

Technology stack summary

How to maintain

automation.

• CI/CD: GitHub Actions, push is deployment
• Monitoring alarms: Prometheus automatically sends emails to mobile phones when exceptions occur
• Data update: scheduled task, automatically pull SPAN parameter every day
• Backup: PostgreSQL Automatic backup to S3, retained for 30 days
• Customer service: 70% of questions are answered in the help center, and the remaining 30% are answered via email and messaging tools

Weekly time allocation:

Have a technical question you want to chat about?hyperstock.werich@gmail.com

Complete blog series

Thanks for reading. From entry to advanced, from data to behind the scenes, I hope these contents will be helpful to you.

Return to blog homepage →

Hyperstock.net — Use AI to make every Sell Put more efficient.