Skip to content
Request access
Plataforma · Gestão de risco

Um enquadramento de risco multicamada para carteiras institucionais. not a report.

Validação antes da transação, métricas de carteira em tempo real, risco de liquidez e um perímetro vivo de limites e interruptores de emergência — calculados a partir dos mesmos dados de posição e de mercado que determinam o P&L, com transparência de ponta a ponta.

The kernel

Five layers of control, one deterministic kernel.

Risk is a perimeter, not a checklist. Each layer hands off to the next against the same live book, and the kernel is the only step that says yes.

  1. Mandate

    The charter. Versioned, human-authored and co-signed. Every change follows propose, four-eyes review, approve, and the author can never approve their own.

  2. Pre-trade

    Checked before it leaves. Every order is evaluated against compiled policy before it reaches a market; an order that breaches a limit is refused, recorded and returned to the agent.

  3. Live book

    Observed continuously. Observers watch exposure, liquidity, volatility and drawdown in real time, and move the kernel between states when thresholds are crossed.

  4. Exit tiers

    Escalating authority. Four tiers of exit, each with its own gate, from a trim the machine may make alone to flattening the book, which only a human may authorize.

  5. State machine

    The firm’s posture. Six states describe what the firm may do right now, from NORMAL to HALT.

Layer II · Pre-trade

The cap is enforced before the trade.

Order-level checks run before anything reaches a market, evaluated against the limit hierarchy compiled from the mandate.

Pre-trade check

Refused, recorded, returned.

  • Proposed weight7.3% single-name
  • Mandate cap7.0% single-name
  • Kernel decisionRefused · returned to the agent
  • LedgerOne row, with the clause that refused it

Illustrative sample data

  • Position and concentrationSingle-name, sector, issuer, geographic and security-type limits, evaluated against the limit hierarchy before an order is submitted.
  • LiquiditySize compared with average daily volume, estimated market impact and time to liquidate, assessed per order.
  • Margin and leverageInitial and maintenance margin, portfolio leverage and gross and net exposure, validated against the mandate’s leverage limits.
  • Mandate restrictionsInvestment-policy adherence and prohibited securities, checked clause by clause. An override is an APPROVAL-GATED exception, never a silent bypass.
Layer III · Live book

Risk, measured as the book moves.

Measures recalculate continuously against the book the firm actually holds, attributable to where the risk sits.

Measured continuously

  • Live limitsAggregate exposure, leverage and drawdown limits, plus sub-portfolio risk budgets, compiled from the mandate and enforced through the session.
  • VaR and expected shortfallHistorical-simulation, parametric and Monte Carlo VaR with expected shortfall, across confidence levels and horizons.
  • Factor exposureMarket, style and macro factor decomposition, so exposure traces to the factors driving it.
  • ConcentrationEffective number of positions, top-N and HHI concentration, and sector, geographic and currency exposure.

Liquidity, profiled

  • Liquidity scoringPosition-level classification and days to liquidate under normal and stressed conditions.
  • Market-impact modelExpected cost of an exit as a function of order size relative to traded volume.
  • Liquidity bucketsHoldings classified into tiers, so concentration in thinly traded names is visible.
  • Redemption stressRedemption and gate scenarios tested against a book’s liquidity profile. On the roadmap.

Stress scenarios, run against the actual book

  • 2008 · Q4 credit shock
  • 2020 · Q1 liquidity shock
  • Rate-taper shock
  • Correlation spike
Layers IV and V · Escalation

Authority escalates with the stakes.

Exits and the firm’s posture each carry their own gate. The machine may trim on its own; only a human may flatten the book or halt the firm.

Exit tiers

  • A · TrimAUTO
  • B · De-riskNOTIFY
  • C · Exit nameAPPROVAL-GATED
  • D · Flatten bookHUMAN-GATED

State machine

  • NORMALAUTOAgents operate inside the full envelope.
  • WARNNOTIFYA soft threshold is crossed. Agents continue; the accountable human is notified.
  • THROTTLEAPPROVAL-GATEDTurnover is capped. New names wait for a named approver.
  • PAUSEDHUMAN-GATEDNo new orders. The existing book is held as it is.
  • FLATTENHUMAN-GATEDThe book is unwound under exit tier D. The kernel may propose it; a human authorizes.
  • HALTHUMAN-GATEDAll activity stops, including agent writes. Restart is four-eyes.

Strategy go-live is APPROVAL-GATED: a strategy is stress-tested against the scenario library before a named approver lets it trade. Every escalation, approval and override is written to the decision ledger.

On the record

Every number, explained.

Risk tooling earns trust by showing its work. Each measure, limit check and breach traces to the inputs and the clause behind it.

  • Attributable. Exposure decomposes by position, sector and factor, so a number always traces to its source.
  • Traceable. Every limit check and breach is a ledger row carrying the clause it was evaluated against. Written once, never edited.
  • Inspectable. Opacity is treated as a defect. Ask Ada why a limit fired and the answer cites the ledger row.
  • Custom models. Firm-specific risk calculations and proprietary models plug into the same framework and the same record.
Six-week shadow pilot

Run the kernel beside your book.

The kernel reads your mandate and shadows your book for six weeks, with no capital at risk. Every check it would have made is on the record for you to inspect.