Quant Capital Engineering
Menu

Signal is not permission

QCE teaches a trade decision as a chain you can inspect: assumptions, evidence, risk controls, permission and recorded outcomes. Each link can be checked, and each can stop the chain.

A signal is input, not authority

A signal, a forecast or a model output is useful: it is how you estimate what might happen. It does not decide whether new exposure is allowed. That decision needs evidence about how good the estimate is, limits on what can be lost, inputs that are current at decision time and an explicit record of who or what granted permission.

None of this is new. Systematic trading, validation and risk control are established disciplines, taught well elsewhere. QCE’s contribution is the connection between them: every course artifact links an assumption to the evidence for it and to the decision it allowed or refused.

  1. AssumptionsWhat is taken as given, written down
  2. EvidenceObservations, experiments and their limits
  3. Risk controlsLimits that can veto new exposure
  4. PermissionA bounded, expiring authorization
Read top to bottom. Any link can stop the chain.

The teaching sequence

Lessons follow eleven steps grouped in four stages. Real systems may combine steps or run controls in parallel; the order is for learning, not a runtime design.

  1. Understand

    1. 01Observe
    2. 02Represent
    3. 03Classify

  2. Estimate

    1. 04Compare
    2. 05Estimate
    3. 06Challenge

  3. Authorize

    1. 07Validate
    2. 08Authorize

  4. Act and learn

    1. 09Execute
    2. 10Measure
    3. 11Re-evaluate

One candidate through the checks

This is the synthetic baseline case from the sample lesson. Its arithmetic passes, and it still does not receive permission, because the evidence check is incomplete.

Diagram grammar: rectangles are steps, a gate mark is a constraint check, rounded shapes are end states. Solid lines were evaluated for this candidate; dashed lines were not.

  1. Candidate F02 baselinep 0.56, gain 12, loss 10, cost 1.80. Teaching policy v0.3.
  2. Inputs fresh at decision time?Yes
    If no (not evaluated for this candidate)
    Refuse new risk (INPUT_STALE)
  3. Estimated value above zero after costs?Yes, +0.52 estimated units
    If no (not evaluated for this candidate)
    Reject (COST_EXCEEDS_ESTIMATED_EDGE)
  4. Evidence and policy checks complete?No
    No (taken for this candidate)
    No permission (EVIDENCE_INCOMPLETE)
  5. Permission issued (teaching example only)

New risk and risk reduction are different actions

Refusing new exposure is not the same as closing what already exists. QCE teaches them as two classes with separate controls.

NEW_RISK

Any action that can increase exposure. It is denied by default and needs explicit, current permission.

A stale input stops fresh exposure, whatever the calculated value says.

RISK_REDUCTION

Reducing or closing known exposure. It follows its own defined controls, and an entry rule must never block it.

A stale input does not justify blindly sending an unverified market exit. Unknown positions are reconciled first and escalated to an operator under a written failure plan.

Permission is not an order

An execution permit is a bounded authorization record for one proposed action. It names the candidate, instrument, side, maximum quantity or notional, price tolerance, issue and expiry times, policy version, evidence reference and a single-use identifier.

Permission can expire or be revoked, and material conditions are rechecked when it is used. Each later state is a separate fact that must be recorded.

  1. PermissionBounded, single use, expires
  2. OrderSent within the permit's limits
  3. FillWhat actually executed
  4. LedgerThe fill recorded and reconciled
A permission is not an order, an order is not a fill, and a fill must reach the ledger.

Plain meanings for QCE terms

Every course term maps to an ordinary engineering concept.

Market state
A versioned set of features available at a stated timestamp.
Regime
An uncertain classification or conditioning rule, not a fact about the market.
Evidence
Recorded observations, experiments and assumptions, including their limits.
Risk authority
A logical component that applies constraints. It is not a regulator and not an intelligent guarantee.
Execution permit
A bounded authorization record for one specific proposed action.

Expected value is an estimate

For the elementary binary example used in Foundations:

EV = pG − (1 − p)L − c

Here p is a probability estimate, G and L are gross gain and loss magnitudes in the same units, and c is total expected round-trip cost. Real systems need full outcome distributions and costs that depend on size.

Expected value is conditional and model-dependent. A positive expectation does not bound any particular loss, prevent ruin or guarantee a positive return over a sample.

Trying to invalidate a result

Validation teaching covers chronology, leakage, repeated experimentation, nonstationarity, dependence, costs and baselines.

  • Selection matters: the number of trials behind a result is recorded, because testing many variants inflates the best one.
  • Chronological folds help but do not by themselves remove overlapping labels or every form of leakage.
  • Ten thousand correlated observations are not ten thousand independent trials.
  • Shuffling trades does not automatically create independent plausible futures; resampling needs an explicit dependence model.

What passing means

Assessment results say Course assessment passed. They never say a system is validated. A profitable-looking chart is not the target; a well-supported rejection is a successful outcome, and a dossier that recommends abandoning a strategy can pass.

A completion record is not accreditation, a professional designation or evidence of trading competence.

References

  1. Bailey, Borwein, López de Prado and Zhu. The Probability of Backtest Overfitting. Journal of Computational Finance 20(4), 2017. Open-access copy (eScholarship)
  2. scikit-learn. TimeSeriesSplit API documentation. scikit-learn.org

See the method applied

Open sample lessonCompare programs