Northgate Capital

Mobile application

The challenge

Northgate's risk numbers were a day old. Positions reconciled overnight in a batch that finished around 4 a.m., meaning portfolio managers traded all day against yesterday's exposure. During the March volatility event, one desk breached its concentration limit at 10 a.m. and nobody knew until the following morning. Month-end close took 34 hours of largely manual reconciliation, and the break log routinely ran past 300 unexplained differences.

Research

• Traced every calculation in the existing risk stack, including 14 spreadsheets that turned out to be load-bearing
• Interviewed portfolio managers, risk officers, and operations staff across all three trading regions
• Analyzed twelve months of reconciliation breaks to classify root causes
• Benchmarked latency requirements against the actual decision cadence on each desk

The solution

We rebuilt position keeping as an append-only event ledger in Go. Every execution, corporate action, and adjustment is an immutable event; positions are a projection that can be recomputed deterministically from any point. This eliminated the entire class of break where the position and its history disagreed. Risk calculations subscribe to the same event stream, so exposure updates within 800ms of a fill rather than overnight. ClickHouse backs the analytical surface, letting risk officers slice twelve months of tick-level history without touching the transactional path.

UX decisions

Features

• Append-only event ledger with deterministic position replay
• Sub-second exposure and VaR across nine asset classes
• Pre-trade and intraday limit checking with escalation workflows
• Performance attribution by factor, sector, desk, and strategy
• Corporate action processing with automated position adjustment
• SOX-traceable regulatory reporting with full calculation lineage
• Client portal exposing mandate-level performance and allocation

Technology

Architecture

Execution feeds from three OMS providers land in Kafka with strict per-instrument ordering. A Go ledger service validates and appends events, then publishes projections to PostgreSQL for transactional reads and ClickHouse for analytics. Risk calculators are stateless consumers that rebuild from the log on deploy, which makes model changes safe to roll out and trivial to roll back — we can replay a full trading day through a new model in under four minutes. Redis fronts the hot position cache. The whole estate is Terraform-managed across two regions with active-passive failover rehearsed monthly.

Results

Position latency settled at 800ms p99 from execution to reflected exposure. Unexplained reconciliation breaks fell 91%, and the residual set is now dominated by genuine upstream data issues rather than internal disagreement. Month-end close compressed from 34 hours to 6. The risk committee approved retiring the legacy batch entirely after two parallel closes — a decision the CRO had expected to take a year of evidence.

Lessons learned

1. 01The load-bearing spreadsheets were the real project. Discovering them took six weeks and reshaped the architecture more than any stated requirement did.
2. 02Parallel running through two full month-end closes was expensive and non-negotiable. The first close surfaced three model discrepancies that no amount of unit testing had caught.
3. 03Deterministic replay turned out to be the feature that sold the platform internally. It was designed for correctness, but its real value was letting risk officers answer 'why did this number change' without calling engineering.