Backtesting benchmark: 31 October 2026
This is the benchmark package behind The fastest backtester you’ve never heard of is built in plain R. It uses the accepted October 31 measurement snapshot.
Open the full HTML report · Download the source and evidence bundle · Download every measured clock
What was measured
A synthetic SMA5/10 crossover rule targets zero or one share per instrument. Books begin flat with USD10,000,000; fees and slippage are zero. Signals use the current close, with intended fills at the next input bar’s open. First usable scores seed the rule, unknown signals hold, and final-bar changes cannot fill. Native execution differences are classified explicitly.
The article’s four rows have five completed repeats per workflow. The broader report contains 235 complete processes and 470 cold/warm clocks, plus seven repetitions the maintainer chose not to run. In its separate N1000 session test, ledgr has three durable, two ephemeral and three spot-FIFO repeats; PMwR, vectorbt and PyBroker have five each. The largest stress geometry has one descriptive observation per workflow. The report labels those counts rather than treating every result as n=5.
The larger peer-suite project remains unfinished. These records support the specified SMA workloads and the displayed workflows, rather than a general framework ranking or completed release certification. The original report’s preliminary banner is preserved. Publication shares the bounded measurements discussed in the article; it does not change the archived report’s original internal-review status or certify the complete suite.
Methods
Warm starts with unchanged, already prepared native input. A fresh book, experiment-specific setup, indicator calculation, execution and the required result materialization are timed. Completed books and precomputed indicator payloads are not reused. Each native framework owns its fills and accounting.
Cold begins at process launch and includes imports, CSV input, reusable native preparation, first-use work and the required outputs. It does not claim an empty operating-system, disk or persistent JIT cache. Installation and common fixture generation are outside both clocks.
The N10/100/1000 session cases execute that many independent experiments per arm. They are measured batches, not multiplication-based projections or one N-candidate grid. Progress/identity checks are inside the batch clock; export, financial verification and explicit between-arm collection are outside.
Durable ledgr and the full-output peers materialize fills, cash, equity and asset holdings. Ledgr’s reduced-output sweeps retain equity and closed-trade P&L and receive no full-output peer ratio. PMwR’s native reported journal prices are distinct from its cash execution prices. All nine PMwR/vectorbt qualification cases agree exactly on cash, equity, holdings, decisions and fill keys/quantities; reported prices do not match. This is portfolio-state agreement, not an assertion of identical reported fill prices.
| Component | Recorded version or policy |
|---|---|
| ledgr | Frozen product build 891c23c8, before the new core |
| PMwR | 1.3-0 |
| vectorbt | 1.2.0; Numba explicitly selected for indicators and simulation |
| R | 4.6.1 |
| Python | 3.12.13 |
| CPU | Intel Core i9-12900K, 16 physical cores / 24 logical processors |
| Installed RAM | 64 GiB |
| OS | Windows 11 Pro, 64-bit, build 26200 |
| Host policy | High performance; process-tree affinity 0xFFFF, first 16 logical processors |
| External math / Numba threads | 1 |
| Observed ledgr DuckDB thread setting | 16; not privately overridden |
The CPU/RAM inventory was captured on November 1, after the measurements; those model details were not in the original timing rows. The timestamp and full values are retained in source.json. Background activity was observed rather than used as a refusal criterion. The report retains timing ranges, actual order and sampled memory; it does not certify a contention-free desktop. Memory peaks are sampled process-tree lower bounds, with missing short phases shown as missing observations.
PMwR/vectorbt ordinary paired windows alternate first position. Other framework observations remain in their original windows. The final ledgr and PyBroker session windows are separate, not counterbalanced. Fifteen older valid PMwR processes remain unpooled historical evidence.
Data and source
- All cold/warm clocks
- Sampled phase memory
- Native agreement classifications
- Headline portfolio differences
- Exact PMwR/vectorbt paired-state checks
- Required versus executed suite coverage
- Source, locks, checker and compact evidence ZIP
- Artifact provenance and integrity values
The ZIP is an inspection bundle from the accepted ledgr benchmark commit e29407b1f26f6806de9390c9005aaea8e521520c. It includes our native recipes, dependency locks, runner/checker, compact records, independent review and the original audited HTML. It does not include installed third-party libraries or the large raw native captures. Regeneration requires the pinned ledgr product and documented native environments. Rendering this website reads only the files in this repository and never runs a backtest.
The source project is blechturm/ledgr. At draft preparation, the accepted benchmark commit was unavailable on the public remote. The download supplies the exact source/evidence snapshot without relying on that Git reference. The served report differs from the original only in three navigation links pointing to this page. Its 19 tables and embedded plot are unchanged.
Record history
Earlier stopped or withdrawn windows remain in the benchmark’s record history. A Python cold-result retention defect caused one window to be withdrawn before corrected measurement. Later unattributed losses of files from TEMP environments stopped imports or launches. The same exact dependency pins were restored outside TEMP and native outputs requalified.
All 31 completed observations from the interrupted scaling window and three completed R session observations were retained with original metadata; only missing cases were run. Failures were not converted into performance clocks, and no completed observation was selected by speed. The final seven unrun repetitions are the explicit maintainer curtailment, not framework failures or resource-censor claims. The report preserves its excluded historical windows and separately retained older PMwR clocks.