Inspiration

Retail investors are told to "stay diversified" and "know your risk", but when they ask a concrete question, such as "what if I put 5% into NVDA, funded from bonds?", they get either a generic disclaimer or a confident AI answer they cannot check. For the AlphaForge AI challenge we wanted the opposite: an assistant whose every number can be traced and whose every sentence has been verified before anyone reads it.

Aplomb is built around one fictional investor, Robert, a 59-year-old healthcare professional with a CAD 600,000 portfolio, a "preserve capital" objective, a 3–5 year horizon and little interest in crypto. Aplomb tests moves against Robert's own rules, nothing is ever executed, and it is always labelled "Simulation only, not advice."

What it does

  • Decide: Describe a move in plain language ("add 5% of NVDA, funded from bonds") or fill in a form. An instant rule check compares the portfolio before and after, then a full analysis runs. Funding options (pro-rata, from cash, from named holdings, new money) can be compared side by side.
  • Rule checks: 11 explicit rules for Robert, including: no single stock above 10%, no sector above 20%, equities at most 55%, bonds and cash at least 35%, volatility at most 9% a year, a 2022-style rate shock costing at most 9%, maximum drawdown at most 18%, and no crypto. Each rule shows its value, threshold, headroom and a status (fit, tension, conflict, or not enough evidence).
  • Portfolio: Historical return, annualized volatility, maximum drawdown, allocation by asset class and country, a 2022 replay stress test, and effective number of positions, all computed on today's weights applied to past prices.
  • Strategies: Standing rules tested on history: rebalancing, regular contributions and a risk guard that de-risks after a drawdown.
  • Replay (time machine): Pick a past date and a move. The analysis sees only the data available on that date, and the outcome is revealed afterwards.
  • World Explorer: A globe of news, hazards, macro releases, energy and geopolitical events tied back to the portfolio's exposure.
  • Evidence and Audit: Every analysis is stored with its evidence, steps, tool calls and verdict, so any answer can be re-opened and inspected.
  • Sources: A passport for every dataset (provider, retrieval date, coverage, licence) plus a list of known data traps and how each is handled or disclosed.
  • Labs: Expert tools such as covariance, factor exposure, funds-inside-funds look-through, sale timing and liquidity, and sensitivity scenarios.

How it works: code computes, AI explains, an auditor verifies

Aplomb is a multi-agent system with strict roles:

  1. Planner (AI): reads the request and profile, and decides which scenarios, calculations and research questions to run.
  2. Researcher (AI): gathers evidence with citations, never conclusions. A page that is not saved as evidence cannot be cited.
  3. Quant (deterministic code): computes every metric. Each number leaves as a fact with its dataset, method and row counts.
  4. Risk officer (deterministic code): evaluates every rule for every scenario, computes the verdict, the funding comparison and the largest size before the first rule conflict. It runs on every decision and the model cannot skip it.
  5. Writer (AI): drafts the explanation using tokens that point to facts, rules and evidence. It receives the ledger but computes nothing and decides no verdict.
  6. Auditor (code + model review): runs checks A01–A14 against the ledger, strikes or repairs unsupported sentences, computes a ledger hash and signs the result. A second LLM reviewer sees only the sentences and the facts they cite.

The result is that the language model never does arithmetic and never decides a verdict, yet the user still gets a readable explanation.

Data

Daily adjusted prices (Yahoo Finance via yfinance, Nov 2021 to Nov 2026), Bank of Canada Valet (FX, rates, yields, CPI), FRED, Statistics Canada release dates, the World Bank, USGS, central-bank press releases and more, over a 31-instrument universe provided by the organizers. Every dataset has a passport page with provider, terms and coverage.

Challenges

  • Honest data: Handling USD/CAD conversion, FX holiday gaps, a market calendar common to Canada and the US, an odd Yahoo "split" on CNQ.TO, and survivorship bias in a 2026-selected universe. Each trap is either handled or disclosed on the Sources page.
  • No hindsight in replays: Macro series have release dates and revisions, so the replay withholds values that were not yet public on the chosen date.
  • Trust in AI text: Making every sentence traceable to a checked figure, rule or cited source, and letting the auditor strike sentences it cannot support.

What we learned

The most useful design decision was to make the AI's job small: plan, research and explain, while deterministic code owns every number and every verdict. That made the system easier to test, easier to audit, and much easier to trust.

What's next

Look-through exposure for ETF holdings, more research adapters (SEC filings, public-attention signals), and more investor profiles with their own rule sets.

Try it: https://d1r325uzu5fe7x.cloudfront.net/decision-desk/ (simulation only, fictional portfolio, not financial advice).

Built With

  • amazon-cloudfront
  • bank-of-canada-valet
  • fred-api
  • large-language-models
  • multi-agent-systems
  • yfinance
Share this project:

Updates

Submission history