Machine-Learning FX Trading Platform: Model Audit → Cloud-Deployed Paper Trading

Customer

Challenge

• Model validity & bias control: Detect and remove look-ahead/data-leakage issues; re-establish clean train/test splits.
• Backtesting realism: Produce auditable portfolio and risk metrics with reproducible runs and clear assumptions.
• Cost-aware design: Select a prediction/trading cadence that balances signal quality and transaction costs, without revealing timing specifics.
• Cloud & broker integration: Provision a fault-tolerant runtime and integrate a broker API for paper execution, with minimal operational burden.
• Data integrity & ops stability: Ensure logging correctness, safe restarts, and recoverability after interruptions—avoiding disclosure of internal runbooks.

Solution

Approach: A compact Marsbridge squad—Quant Lead, ML Engineer, Cloud/DevOps Engineer, and Research Analyst—delivered iterative drops from model audit to stable paper trading.

Model audit & bias removal

Replicated baseline results; identified leakage/label drift; refactored pipelines to enforce causal ordering and proper evaluation windows.

Backtesting & MLOps

Implemented a vectorized backtester with portfolio-level analytics and added process supervision, health checks, and periodic automated restarts for MLOps resilience.

Paper trading & broker/API integration

Deployed a cloud runtime with broker API connectivity (paper environment), synchronized positions/orders, and implemented defensive order templates.

Technologies & tools

Process

1. Discovery & replication: Confirm scope, reproduce baseline results, and map potential leakage paths.
2. Bias control & refactor: Enforce causal data flow; re-baseline the model on corrected pipelines.
3. Research iteration: Broaden features and targets; run controlled experiment batches with consistent reporting.
4. Backtesting & analytics: Deliver portfolio-level analytics and governance-ready summaries.
5. Paper-trading deployment: Bring up cloud runtime; connect broker API; validate end-to-end orders in a safe environment.
6. Hardening & runbooks: Add observability, automated restarts, and recovery protocols; finalize handover materials.

Results

• Credible baselines: Eliminated leakage; established auditable backtests and clear performance accounting.
• Operational readiness: Stable paper-trading workflows with recoverable processes and standardized runbooks (details withheld).
• Decision clarity: Comparable experiment results and governance-grade reports simplified go/no-go decisions for live staging.
• Future-proofing: A modular path to live trading with configurable timing/parameters—kept private to the client.

All strategy parameters, trade horizons, instrument lists, and PnL snapshots have been intentionally removed or generalized to protect client confidentiality.