wegamekinglc / Derivatives-Algorithms-Lib / 36025160657

push

github

perf: reduce LSMC replay and observation overhead (#424)

* Reuse LSMC replay state and prune dead observations

Out-of-sample hard pricing reconstructed evaluators and recording rows for every batch. Reuse worker-local replay state and borrow the prepared compiled program so the frozen policy can be priced without per-date hard recording buffers.

Prune only future model observation outputs removed by LSMC liveness. Keep all sample dates and resolve historical fixings first to preserve Sobol dimensions, path evolution, and validation contracts. Add parity, ordering, and benchmark coverage for hard/AAD and Black-Scholes/Dupire.

* Simplify LSMC replay and profile control flow

Split the replay profile parser and observation compaction into focused helpers so Codacy complexity checks remain within the repository threshold. Keep the replay behavior and benchmark outputs unchanged.

* Preserve generic compiled MC code generation

Keep PreparedScript_::Compile on its original direct-copy path while LSMC borrows CompiledProgram. The forwarding implementation enlarged the ordinary MC entry point and correlated with reproducible vanilla double compiled slowdowns on the CI runner.

107 of 108 new or added lines in 5 files covered. (99.07%)

15580 of 16976 relevant lines covered (91.78%)

3386821.6 hits per line