Advanced F · Institutional trading
Manage strategies through capital constraints, research governance, and reconcilable attribution.
- Chapter 30 · Paper tradingWhat must Paper, Shadow, and Live Small each validate?
- Chapter 34 · PnL attributionHow can PnL be split into direction, signal, Carry, Funding, Basis, execution, fees, slippage, and residual?
- Chapter 35 · Trading journalHow can journal data measure decision and execution deviations?
- Chapter 36 · Trading systemWhich gates does a Research Pipeline pass from idea to production?
- Trader
- Senior Trader
- Portfolio Trader
- PM
Strategy Portfolio and Capital Allocation
Institutions manage strategies sharing capital and responsibility. List return mechanisms, factors, funding, exits, and lifecycle before allocation. Notional, risk budgets, and available cash are three distinct constraints, none replacing the others.
List Trend, Carry, Relative Value, Flow, and Event with approved capital/losses, venues, and dependencies. Allocations plus cash equal total capital; simultaneous maximum use cannot double-allocate cash. Unapproved candidates consume research resources, not production risk.
Capacity: scale changes strategies
Capacity is executable scale within costs, risk, and deadlines, not total market volume. Books, participation, borrow, venue capital, and crowded exits constrain it. Weak net edges may vanish at low participation.
Compare small/medium/large targets with identical signals and stressed depth. Without validated impact, report executable ranges and unverified items, not precise extrapolated returns. Submit normal and stress-exit capacity, maximum funding use, and downsizing triggers.
Strategy Correlation: shared failure is dependence too
Return correlation is one perspective. Uncorrelated strategies may share feeds, venues, collateral, or execution and fail together. Include statistical dependence and concentration.
Remove the largest venue or primary feed; identify strategies unable to value, exit, or do anything beyond reductions. Assign actions/owners to shared dependencies. Different strategies alone prove no diversification. See Advanced D for factors/stress correlations.
Strategy Decay: separate noise from structural change
Decay may raise forecast errors/costs or lower opportunity frequency/capital efficiency. One loss proves no decay; winning streaks excuse no data drift. Thresholds derive from train/validation distributions and account for repeated-check false alarms.
Decompose net returns into gross signals, trade counts, costs, and remainders. Investigate execution if costs change first, data/mechanisms if features shift. Retain alerts, investigations, reductions, pauses, and restarts. Do not extend observation indefinitely to avoid accepting failure.
PnL Attribution: how to avoid double-counting nine components
Freeze denomination, valuation times, benchmark positions, and price bridges. These teaching definitions do not independently identify causal Alpha. Divide changing positions into fixed intervals; separately listed Basis has already been stripped from prices.
| Component | Definition | Avoid duplication |
|---|---|---|
| Directional | Advance benchmark × decision-to-valuation price change | Not assigned again to Signal |
| Signal | Actual minus benchmark × same change | Active accounting contribution is not predictive ability |
| Carry | Independent holding flows, such as lending interest | Excludes separate Funding/Basis |
| Funding | Funding received or paid | Not also Carry |
| Basis | Independently valued basis change | Not also underlying price change |
| Execution | Actual position × decision minus arrival | Timing impact versus decision only |
| Fee | All actual fees, negative | Not also Slippage |
| Slippage | Actual position × arrival minus fill | Includes spread/impact, not deducted separately |
| Residual | Independent net PnL minus eight components | Retain discrepancies, not Alpha |
Teaching example: actual 2 units, benchmark 1; decision 1,000, arrival 995, fill 997, valuation 1,050; Carry +5, Funding −8, Basis +12, fees 6, observed net $112. Components are 50,50,5,−8,12,10,−6,−4,3, totaling 112. Residual 3 needs investigation, not extra Signal.
Deliberately count Funding in Carry too and inspect differences, then restore definitions. Independent cash/valuation ledgers must provide observed net PnL; summing eight components is not an independent observation guaranteeing zero residual. Cross-currency, options, and complex flows need validated expanded valuation, not blind fixed-spot formulas.
Research Pipeline: failures are assets too
Idea includes observations/counterexamples; Research adds mechanism/data protocol; Testing freezes code/costs/splits; Validated passes economic logic, statistics, OOS, robustness, cost, risk, execution, and regime. Paper/Shadow compare intent, quotes, and achievable outcomes; attractive returns cannot bypass data/risk gaps.
Move one candidate toward paper testing, recording evidence, owners, and rejections per promotion. Retain all attempts, data/code versions, parameters, and dates for independent recalculation. Failed candidates enter a failure library, preventing repeated renamed searches.
Strategy Governance: transitions require evidence
Lifecycle: Idea → Research → Testing → Validated → Paper → Shadow → Live Small → Production → Scaled. Any stage may move to Degraded, Paused, or Retired. This course uses virtual ledgers; later states rehearse governance without real trading permission.
Record old/new states, reasons, evidence, limits, approvers, and rollback. Code, costs, or signal repairs may invalidate prior tests, requiring retesting. A version bump cannot preserve obsolete approval. Pause differs from retirement; restart does not automatically resume old orders.
Pause a Paper strategy for missing data, repair, and return to Testing. Retire another after mechanism failure. Find missing evidence blocking promotion. A single person must not propose, approve, and enlarge risk alone.
Risk Committee: define authority before meetings
Research explains mechanisms, trading explains execution, independent risk inspects exposures/scenarios, and portfolio management approves capital. Small teams may combine roles, but the proposer cannot be the only effective approver. Peers/reviewers can simulate independence without inventing institutions.
Record candidates, objections, gaps, decisions, and next reviews. Unvalued assets, inconsistent ledgers, or hard breaches let risk pause additions without waiting for profits. Submit minutes containing rejection or conditional approval, not universal agreement alone.
Kill Switch: positions remain after shutdown
Manual Kill, Risk Kill, Data Kill, Execution Kill, and Model Kill have separate triggers. Stopping new orders, canceling in-flight orders, and managing positions are separate actions. Unconfirmed cancellations and unavailable venues leave risk. “Stop” does not default to automatic market liquidation.
| Trigger | Immediate action | Evidence before restart |
|---|---|---|
| Manual detection of unmodeled incident | Pause additions and record reason | Joint risk/execution review |
| Loss/exposure breach | Lock additions; apply advance reduction plan | Reapproved limits, balances, remainder risk |
| Stale feeds/sequence gaps | Stop affected strategies | Repair, replay, freshness |
| Order-report timeout | Mark unknown; query/reconcile; forbid blind resend | Orders/fills/positions agree |
| Out-of-range model outputs | Reject; use approved fallback or stay paused | Repair, regression, independent approval |
Rehearse stale data and unknown orders separately, preserving timelines and uncanceled remainders. Stops cannot rely on voluntary model compliance; one normal response cannot restart automatically. Show triggers, states, owners, residual risk, reviews, and authorization.
Project: institutional strategy review package
Submit capital/capacity, dependencies, decay monitoring, nine-component attribution, research registry, lifecycle/committee minutes, and two shutdown drills. Strategy IDs/versions agree everywhere. Follow the Institutional handbook for a $1,000,000 virtual portfolio, assessed on reviewable decisions and survival.
Evidence boundaries of Paper, Shadow, and Live Small
Paper validates intent, ledgers, and rules. Shadow records proposed orders beside actually visible quotes and estimates feasible outcomes. Live Small examines real small-scale fills, costs, and operations. Virtual course exercises cannot infer real execution passed from simulation returns.
Separate expected signals, visible quotes, executable quantity, and fees per stage to find the first data/logic/fill mismatch. Predefine deviations and sample requirements. Breaches return to Testing. Without real fills, mark Live Small unverified rather than substituting simulation.
Find decision biases in journals
Keep advance probabilities, risks, failures, actual actions, and deviation reasons. Separate planned losses from profitable rule-breaking. Winner-only samples confuse discipline and results. Calibration may compare observed frequencies within probability bins or mean squared forecast errors; sparse samples require denominators.
Blindly assess advance information before revealing outcomes. Count predefined chasing, stop-moving, canceled-exit, and overtrading labels. Submit counts, context, and testable next actions, not personality judgments from one error. Daily reviews address risk, weekly habits/expectancy, monthly factors/risk-adjusted returns, and quarterly capital/research.