QUESTIONS INTO EVIDENCE.

QUESTION

Where does lookahead bias actually enter quantitative pipelines — and can it be prevented structurally?

HYPOTHESIS

Lookahead bias enters through data-layer engineering decisions, not strategy logic, and can be eliminated by point-in-time access semantics.

METHOD

Audit of ingestion, normalization, resampling and join paths; adversarial tests that attempt to extract future information through every access pattern.

EVIDENCE

Adversarial test suite in quant-market-data-replay — deliberate future-information extraction attempts fail against point-in-time reads.

RESULT

Point-in-time safe reads enforced in the data layer prevent lookahead by construction. Consumers cannot opt out.

LIMITATION

Consumer code can still misuse resampled windows above the data layer; the guarantee covers access, not strategy logic.

QUESTION

Does determinism in the backtest harness measurably reduce false-positive strategies?

HYPOTHESIS

If identical inputs produce identical outputs, parameters tuned to noise become identifiable as unstable rather than performant.

METHOD

Deterministic replay of the same strategy configuration across repeated runs; comparison of results under perturbed data ordering with fixed tie-breaking rules.

EVIDENCE

Apex Quant Engine backtests reproduce byte-identically from raw bars to equity curve; 5,867-trade master-fund runs regenerate exactly.

RESULT

Deterministic harnesses make an entire class of irreproducible 'results' impossible to mistake for findings.

LIMITATION

Determinism addresses reproducibility, not statistical overfitting to a single historical window — that requires out-of-sample discipline.

QUESTION

Where does binary floating-point representation corrupt market data, and does it matter?

HYPOTHESIS

Float representation of prices and monetary values accumulates representation error that surfaces in aggregates, joins and risk calculations.

METHOD

Comparison of ingestion-storage-replay paths under exact decimal arithmetic versus IEEE-754 floats, tracking divergence through aggregation.

EVIDENCE

quant-market-data-replay stores and replays with exact decimal arithmetic across the full pipeline.

RESULT

Representation drift is eliminated as a failure mode; joins and aggregates behave identically across runs.

LIMITATION

Exact arithmetic trades throughput for correctness — appropriate at research speeds, costly at HFT scale.

QUESTION

What changes when risk controls run before order construction rather than as post-trade reconciliation?

HYPOTHESIS

Structural pre-trade controls — risk cap, R:R floor, cost-budget filter — change the population of admitted trades, not just their outcomes.

METHOD

Pipeline architecture with controls as gates: position size derives from stop distance and risk cap; setups failing the 1:4 R:R floor or cost budget never reach execution.

EVIDENCE

automated_trading_os implements 1% per-trade risk cap, minimum 1:4 R:R, fee/slippage modelling, cost-budget filtering, profit lock and giveback limit.

RESULT

The admitted-trade population is defined by rules rather than discretion; every admitted trade satisfies all structural constraints by construction.

LIMITATION

Ablation studies quantifying each control's individual contribution are ongoing.

QUESTION

What is the character of windows where higher, mid and lower timeframes agree?

HYPOTHESIS

One-way information flow — LTF triggers inside MTF structure within HTF direction — produces a distinct, analyzable population of setups.

METHOD

Layered pipeline analysis: HTF establishes directional context, MTF maps structure, LTF times entries; agreement windows extracted and studied.

EVIDENCE

automated_trading_os pipeline with explicit stage contracts; timeframe explorer instrument in development.

RESULT

In progress — agreement-window extraction across instrument history under construction.

LIMITATION

Results not yet published; the pipeline exists, the study is ongoing.

QUESTION

How much of a multi-strategy portfolio's simulated performance is regime selection rather than strategy skill?

HYPOTHESIS

Strategy performance clusters by market regime more strongly than by strategy family — allocation across regimes matters more than strategy count.

METHOD

Decomposition of master-fund backtest results by regime classification across the 11.9M+ bar history.

EVIDENCE

Apex master-fund backtest artifacts: +570.18% net simulated ROI across 5,867 completed backtest trades.

RESULT

In progress — regime attribution analysis under construction.

LIMITATION

Regime classification itself is a modeling choice; attribution results will depend on the classifier.

QUESTION

How should equal-timestamp market events be ordered so replay is reproducible?

HYPOTHESIS

Without deterministic tie-breaking, equal-timestamp events order arbitrarily across runs, making replay irreproducible.

METHOD

Reorder stage with explicit deterministic tie-breaking rules; repeated replay verification that output is byte-identical.

EVIDENCE

quant-market-data-replay determinism verification: identical replays across runs including tie-breaking.

RESULT

Replay determinism holds under nanosecond timestamps with equal-timestamp events.

LIMITATION

Tie-breaking rules encode an ordering assumption — an approximation of true within-timestamp causality.

QUESTION

What evidence threshold should graduate a strategy from research to deployment consideration?

HYPOTHESIS

A governance gate — verified artifacts, reproducible runs, documented limitations — raises the evidentiary floor of everything promoted.

METHOD

QRSIP workflow: hypothesis → experiment → verification → artifact → report → promotion, with 352 tests on deterministic fixtures and CI security validation.

EVIDENCE

QRSIP test suite and promotion-gate integrity tests; alpha-stage platform operational for the core workflow.

RESULT

Strategies cannot pass promotion without verified evidence; the gate is structural.

LIMITATION

The gate verifies process integrity, not edge — a well-governed strategy can still be unprofitable.