3
live funds monitored
Tools
Tools
Loading
Tools
Tools / Asset Management / Risk
Risk-AM records the transactions of three student-managed portfolios, monitors their risk and gives each fund team an auditable basis for its portfolio discussions.
3
live funds monitored
7
task workspaces
918
automated tests passing
5
Monte Carlo methods
Use the platform
This is the public operating interface, already open and interactive. Follow the three steps below, or enter the full platform in a new tab.
Choose a fund
Start with Guardian, MarketPlus or Horizon.
Open an analysis
Move from overview to risk, performance or group exposure.
Inspect the evidence
Read freshness, methodology and model limitations in context.
Starting the interactive platform
Connecting to the public Risk-AM interface.
A working system
Risk-AM replaced fragmented portfolio work with a reproducible operating layer for Guardian, MarketPlus and Horizon. The same system records activity, computes risk, checks mandates, supports research and exports the sanitised data used by TIC's public strategy pages.
University students at Instituto Superior Técnico designed and built it, and the Risk team now uses it in its weekly work. The platform has 918 automated tests, a quality and security workflow that blocks a bad run, and a deployment configurator for a different set of funds, branding and providers.
Frontier methods where they add value. Deterministic controls where the numbers must hold.
System state
Risk-AM · release 2026.07
Supabase Postgres
Transactions, members and audit events, with a SQLite fallback
Compute once, read many
Heavy work runs nightly; the dashboard reads the snapshot
Chained providers
CSV, then a licensed vendor, then Yahoo, plus FRED for rates
Role-aware
Member, PM and package entitlements, enforced before routing
One platform, six layers
The product connects quantitative finance, software engineering and team governance. Each layer can be inspected independently; none depends on an LLM to calculate portfolio risk.
01
Historical and parametric VaR and CVaR, cross-checked against EWMA and GARCH volatility. Drawdown depth, duration and recovery, liquidity and days-to-liquidate, beta, tracking error and Fama-French five-factor exposure.
02
Five Monte Carlo methods: geometric Brownian motion, Student-t fat tails, historical bootstrap, block bootstrap and Ledoit-Wolf. Plus historical stress scenarios and a scenario builder you drive yourself.
03
VaR decomposition with marginal and component contributions, Kupiec backtesting of the model itself, and an efficient frontier built on Ledoit-Wolf covariance shrinkage rather than a raw sample matrix.
04
Brinson-Fachler allocation and selection, Carino multi-period smoothing, Karnosky-Singer currency decomposition, and explicit residual reconciliation. Where an ETF look-through is missing, it is disclosed rather than assumed.
05
Mandate limits read from one config file, pre-trade compliance and execution-cost estimates, and data-quality checks severe enough to block publication outright. Every run leaves a manifest and an audit trail.
06
A tutor wired into the KPI you are looking at: 33 reviewed concepts, 16 lessons, 18 methodology guides and a seven-level path with prerequisites and a capstone. No formula or number in it is LLM-generated.
Controlled intelligence
Portfolio numbers come from Python and stored analytical outputs. LLM features receive a bounded evidence set, use task-specific inference settings and are clearly separated from numerical controls.
Deterministic lexical RAG selects relevant portfolio sections without an embeddings service. The default production chain runs OpenAI's GPT-OSS models through Groq. OpenAI and Anthropic adapters activate only when configured; NVIDIA is isolated to an explicit public-data evaluation route.
Deployment configuration · 24 July 2026
Output target is the provider-neutral profile setting. The GPT-OSS completion budget also covers hidden reasoning tokens.
Profile 01
Numerical Q&A
Profile 02
Reports & scenarios
Profile 03
Specialist analysis
Profile 04
Validated JSON & repair
Profile 05
Final synthesis
Retrieval
Relevant evidence only
Context
Six recent messages, bounded
Public evaluation
NVIDIA DeepSeek V4 Pro → Groq GPT-OSS 120B; public data only
Boundary
Research support, never the risk engine
Architecture
Heavy analytics run outside the interface and land in versioned artefacts. The dashboard reads stable outputs instead of rebuilding portfolios on every visit; GitHub Actions orchestrates the recurring workflow.
Trades, members, watchlists and audit events live in Supabase Postgres.
A Python pipeline replays portfolios and runs the analytical engines.
Data-quality, mandate and operational controls gate the output.
SQLite snapshots and Parquet aggregates preserve reproducible results.
Cloudflare R2 serves versioned artefacts to the deployed application.
Streamlit powers the team platform; sanitised exports feed the public site.
Built inside Asset Management
Risk-AM belongs to the Risk Team. The tools catalogue is the framework through which future department-built systems can earn the same visibility.
Student-built research and operating platform. Indicative market data may be delayed or incomplete. Not investment advice.