Architecture

Microalpha is event-driven so timing assumptions remain explicit and testable.

Event lifecycle

sequenceDiagram
    participant D as Data handler
    participant E as Engine clock
    participant S as Strategy
    participant P as Portfolio
    participant X as Execution planner
    participant B as Broker/materializer

    D->>E: MarketEvent(t)
    E->>P: mark current state at t
    E->>B: materialize plans due at t
    B->>P: FillEvent(t)
    E->>S: observe state and market at t
    S->>P: SignalEvent(t)
    P->>X: OrderEvent(t)
    X-->>E: ExecutionPlan(t+1, qty)
    Note over E,P: No future price or state mutation occurs at t

The built-in market-data executors plan timestamps and quantities without reading future prices or volumes. When the matching symbol and timestamp arrive, the broker materializes the slice using information available at that event. TWAP and implementation-shortfall schedules are fixed ex ante. The safe VWAP path uses equal ex-ante slices until an explicit historical volume profile is provided; it never sizes from realized future volume.

Signals may specify target_weight, which the portfolio translates into the delta between current and desired notional. Repeating the same target therefore does not create a second full-size order; resizing, closing, and flipping remain explicit. The legacy weight field retains its historical full-order sizing behavior for compatibility.

Component boundaries

Component	Owns	Must not own
Data handler	ordered observations and availability metadata	strategy selection
Strategy	signals at the current event	fills or future prices
Portfolio	sizing, cash, positions, exposure, turnover, risk	market-data revision logic
Execution planner	timestamps and slice quantities	future observed price/volume
Broker/materializer	fill at a due event	model selection
Walk-forward evaluator	train/test/holdout isolation	post-hoc strategy mutation
Evidence layer	schemas, hashes, reports, claim gate	reconstructing missing facts from prose

Statistical control

centered_max_statistic_test accepts an aligned candidate-return matrix and an explicit benchmark series. It computes candidate-minus-benchmark statistics, recenters all differentials under the null, and synchronously resamples rows. This is the selection correction used by Audit Lab, the public Market Risk Case, and walk-forward grid evaluation.

The older relative "best versus other candidates" SPA interpretation is not a claim that any model beats a benchmark. Public claims use the explicit benchmark-differential test.