QuantBT Benchmarks

Phase 7 introduces a reproducible benchmark harness for the upgraded backtest backends.

python3 benchmarks/run_phase7.py --profile smoke
python3 benchmarks/run_phase7.py --profile standard --repeats 5
python3 benchmarks/run_phase7.py --profile standard --repeats 5 --no-tracemalloc
python3 benchmarks/profile_phase7.py --profile standard --repeats 3

Profiles:

• smoke: quick local sanity check.
• standard: commit-to-commit comparison target.
• large: stress profile for optimization decisions.

The runner writes both JSON and Markdown into benchmarks/out/ by default. Nautilus is optional and skipped unless --include-nautilus is passed.

Backends currently measured:

• native_vectorized
• native_event
• native_event_prepared
• portfolio_legacy
• native_portfolio
• optional nautilus

The committed summary lives in benchmarks/phase7_report.md. Local JSON/MD outputs under benchmarks/out/ are git-ignored by design.

When a backend misses a runtime threshold, run profile_phase7.py before considering Cython/C++. The committed profiling summary lives in benchmarks/phase7_profile_report.md.

Phase 9 optimization follow-up:

• benchmarks/compare_phase9_parity.py checks that optimized sizing/order compilation does not change target units, equity, positions, order reports, or fills.
• benchmarks/phase9_optimization_report.md records the first post-profiling optimization pass and remaining bottlenecks.
• native_event_prepared measures the WFO/service pattern where market arrays and compiled order arrays are prepared once and replayed through the same event/accounting kernel.
• --no-tracemalloc is available when comparing runtime separately from memory instrumentation overhead. Use the default traced mode when peak memory is the metric under review.

Phase 14/16 service-loop follow-up:

python3 benchmarks/run_phase14_service_loop.py --rows 1440 --symbols 6 --trials 8 --repeats 2
python3 benchmarks/run_phase16_performance_debt.py --rows 1440 --symbols 6 --replays 8 --repeats 2

• phase14_service_loop.* decomposes WFO, native-event, arbitrage and report workload costs.
• phase16_performance_debt.* compares normal endpoint replays with endpoint.prepare_service_context(...) and records the current Cython/C++ decision.

Phase 49B WFO prepared/scalar certification:

python3 benchmarks/run_phase49b_wfo_performance.py --rows 1000 --trials 16

• compares Phase 49A reference retention with Phase 49B prepared context, scalar trial scoring and compact ledgers using identical mathematical work;
• checks exact equity, positions, selected params, objectives, trial order and candidate order;
• separates warm runtime from isolated child-process RSS and records strategy, scorer, market preparation, signal packing and metric-report timing;
• does not cache arbitrary strategy indicators or signal output.

Options Phase 10:

python3 benchmarks/run_options_engine.py --snapshots 96 --contracts 48 --packages 96 --repeats 3
python3 benchmarks/gamma_scalping_backtestsample.py --snapshots 90 --seed 42
python3 benchmarks/gamma_scalping_backtestsample.py \
  --real-options-csv /root/bobby/pool_alpha/alphas_storage/option_based/options_full_history.csv.gz \
  --underlying-source spot \
  --hedge-timeframe 1h

• options_phase10_baseline.* records prepared-tape and compiled-package cache parity for the native option backend.
• The benchmark reports snapshots, contracts, quotes, packages, fills, hedges, memory, uncached runtime, cached runtime, and run-manifest hashes.
• gamma_scalping_backtestsample.py is a runnable long-straddle gamma-scalping smoke sample. It keeps the original research helpers, then runs the public QuantBTEndpoint.options(...) path through build_gamma_scalping_strategy_run(...), strategy_run, underlying, and prepared-cache parity.
• The real-data mode converts legacy Binance options CSV history into QuantBT's canonical option-chain schema, selects an ATM call/put pair with entry/exit quotes, and loads BTCUSDT spot or USD-M perpetual candles from _get_data for first-class delta-hedged combined-equity accounting.
• Cython/C++ should only be considered after a larger profile shows pure kernels, not pandas/tape/report facade work, dominating runtime.

Phase 47C Grid 2,000-bar parity and RSS:

MPLCONFIGDIR=/tmp PYTHONPATH=/root/bobby/pool_alpha \
poetry run python benchmarks/native_event/benchmark_grid_2000.py \
  --grid-module-dir /root/bobby/pool_alpha/alphas_storage/TA \
  --backend python --mode scalar --grid-mode long_only --bars 2000

MPLCONFIGDIR=/tmp PYTHONPATH=/root/bobby/pool_alpha \
poetry run python benchmarks/native_event/benchmark_grid_2000.py \
  --grid-module-dir /root/bobby/pool_alpha/alphas_storage/TA \
  --backend rust --mode audit --grid-mode long_short --bars 2000

The runner uses one warm-up and five measured runs in a backend-isolated process and writes JSON with command/audit fingerprint, terminal accounting, runtime, CPU time, peak/post RSS, and repeated-run RSS slope. See docs/grid_native_event_phase47c.md for the parity contract and backend policy.

Phase 31 intrabar execution:

python3 benchmarks/run_phase31_intrabar.py --rows 25000 --repeats 3
python3 benchmarks/run_phase31_intrabar.py --rows 512 --repeats 1

• phase31_intrabar_benchmark.* compares the new fast intrabar_bracket_v1 kernel against the close-target pure kernel, the Python intrabar oracle, fill replay, and the native-event explicit-order facade.
• Use the fast intrabar route for single-symbol next-open SL/TP/trailing research. Use report_level="audit" for fill-ledger certification and report_level="minimal" for WFO/optimizer loops.