The shared backtest core #

A backtest is a controlled replay. A strategy sees a historical information set, makes a decision, and receives a simulated outcome. The dangerous part is not writing a trading rule. It is making sure the replay did not peek, change assumptions between variants, or reward itself with an unrealistic benchmark.

The shared backtest core puts those responsibilities in one public evaluator. A study supplies the question. The core supplies the measurement rules.

The strategy specification #

A strategy specification is a frozen description of how decisions become simulated positions. It names whether the study reacts to events or rebalances a portfolio, its decision time, fill point, exit rule, notional or capital basis, order priority, dividend treatment, universe, and liquidity rule.

The decision function must be pure and importable. Mutable closures and parameters that cannot be serialized are refused. That makes the run identity reproducible and lets serial and parallel runs produce the same bytes.

The core supports later-session opens, declared auction or intraday points, conditional exits, and portfolio targets. A fill must occur after the information used to decide it. If a study uses an official open as a pre-market indication, the report labels that approximation rather than hiding it.

Every event study must also say what happens when its intended exit lies beyond the evaluation window. window_end="force_close", the compatibility default, closes at the last evaluable point; window_end="unevaluable" excludes and counts the incomplete outcome. The separate require_complete_path declaration decides whether a missing or invalid held-session bar makes the whole trade unevaluable. The defaults preserve earlier result bytes, but studies must declare both choices rather than inherit them silently.

No-peeking data views #

Each input declares when its fields became available. The view also has a hard maximum date. Asking for a close, volume, later bar, future membership row, or holdout row too early raises an error. The core does not drop the offending row to make a run pass.

Listing intervals and trailing liquidity determine eligibility when that history exists. A source that contains only current listings is marked SURVIVOR-BIASED SOURCE. Delisted positions exit at the last available close when possible; otherwise the study applies and counts its predeclared delisting fallback.

Costs and fills #

Every run selects one named primary cost profile and at least one distinct harsher sensitivity. Costs apply per side. Liquidity tiers, spread, adverse movement, fees, participation, and funding remain separate components instead of one unexplained haircut.

The core can use fractional shares, so small capital does not gain or lose an arbitrary advantage from whole-share rounding. It reports position notional relative to trailing dollar volume and, when supplied, opening-auction volume. Missing bars remain missing.

An optional independent price source recomputes the trade result only where that second source has coverage. The report shows signed price differences and uncovered trades. It never fills a gap in the second source with the first source and calls that agreement.

The benchmark rule #

The normal primary comparison is:

strategy net return minus benchmark gross return over exactly the same window

The benchmark can be SPY, cash, or the equal-weight eligible universe. The report also checks that the strategy’s own net return is positive. A cost-bearing benchmark exists only for an explicitly labelled allocation comparison, where both sides pay their declared costs.

This distinction prevents a strategy from beating an unfairly costed benchmark while losing money itself.

Statistics that share one clock #

The evaluator reports trade-level mean and uncertainty, full calendar-session returns, annualized Sharpe, DSR with a required trial-census input, stationary-bootstrap intervals, drawdown, worst month, exposure, turnover, folds, recency, and capacity.

Event strategies include zero on eligible days with no position. Capital is based on the maximum concurrent slots times slot notional. These choices stop rare trading from manufacturing an impressive Sharpe by changing the denominator or removing quiet days.

One shared price panel #

The runner builds one immutable dense columnar price panel for a study, with constant-time ticker and session coordinates. Forked Linux workers inherit the same read-only pages, so they do not each serialize or rebuild a market-sized frame. Point-in-time checks, order precedence, fills, costs, and ledger serialization remain on the same path as before the optimization.

The registered synthetic benchmark measured the change on 2026-10-02:

• On 3,000 stocks × 3,800 sessions, 16-worker simulation fell from 989.05 seconds to 0.847 seconds (1,167.5×). The legacy one-core run was stopped after more than 55 minutes; optimized simulation took 7.75 seconds after a separate 133.41-second panel build.
• On the reduced 600-stock × 3,800-session profile, serial simulation fell from 1,557.73 seconds to 6.69 seconds (232.8×), while 16 workers fell from 124.14 seconds to 0.558 seconds (222.5×).
• Peak worker memory fell from about 6.85 GB to 0.82 GB. Peak parent memory fell from 7.21 GB to 1.67 GB, and nominal parent-plus-worker memory fell from 116.83 GB to 14.76 GB.

Performance did not buy a different answer: all 295 captured pre-existing JSON outputs, totaling 53,391,323 bytes, matched byte for byte, including the power, size, and benchmark-strategy runs.

A synthetic proving ground #

The evaluator tests itself on generated markets where the truth is known. The synthetic world contains fat-tailed daily and overnight returns, liquidity tiers, delistings, a planted tier-specific gap-down effect, cross-sectional momentum, and pure noise.

The final registered proof detected the planted edge in 48 of 50 seeded runs. On 200 pure-noise runs it rejected 7 times, inside the exact 99% binomial band of 3 to 19. Six canaries covering look-ahead, survivorship, costs, benchmarks, native simulation, and independent-price behavior passed. Serial and parallel reports were byte-identical.

That result proves the evaluator can detect this known synthetic effect at the registered size. It does not prove that any real strategy works. Synthetic power is a test of the ruler, not the thing being measured.