wegamekinglc / Derivatives-Algorithms-Lib / 35496610591

92%

Build Type

push

github

Committed by web-flow

Commit Message

DAL-283: fuzzy AAD early-exercise valuation (T4) (#401)

* Add fuzzy AAD valuation for early exercise (T4)

EXERCISE products now value in AAD mode: the frozen hard-decision policy
from the double phases feeds a fuzzy replay in which each worker records
payments, exercise values, and condition degrees on its own tape and prices
paths with the recursive blend V_k = d_k*h_k + (1-d_k)*(p_k + D*V_{k+1}),
d_k = CSpr(h_k - C_k(z_k), eps) * condition degree. Coefficients stay
passive tape constants (N6 envelope), and adjoint reduction follows the
thread-count independent batch layout, so AAD PV and risks are bitwise
thread invariant (N9).

- compiler: lift the fuzzy EXERCISE rejection; append LsmcFuzzyPays/
  LsmcFuzzyPaysConst/LsmcFuzzyExercise opcodes recording into the typed
  LsmcFuzzySinks_; thread the Lsmc_ dispatch flag through If/IfElse/FuzzyIf
  branch evaluation so recording opcodes inside branches resolve (the hard
  IfElse case was latently unreachable for LSMC streams too)
- tree: FuzzyEvaluator_ gains optional LSMC recording sinks; without them
  the plain fuzzy arithmetic is untouched
- driver: MCLsmcAadSimulation reuses the double Phases A/B and routes from
  MCAADSimulation; RegressionPredict becomes a template shared by both
- tests: AAD vs policy-regenerating central differences (N6 exceedance
  attributed and recorded, guarded at the measured envelope scale),
  smooth-to-zero convergence band/trend, AAD bitwise thread invariance,
  historical FIX x EXERCISE analytic risks (snapshot and global store),
  and tree/compiled AAD parity

Co-authored-by: multica-agent <github@multica.ai>

* Split the fuzzy replay batch into workspace and helpers

Codacy flags RunFuzzyReplayBatch with cyclomatic complexity 11 (limit 8):
the batch setup, both engine-specific path evaluations, and the replay
loop lived in one function. Extract the per-worker state into
FuzzyReplayWorkspace_ and the tree/compiled path evaluations plus the
shared replay loop into nam... (continued)

Coverage Stats

198 of 203 new or added lines in 6 files covered. (97.54%)

52 existing lines in 6 files now uncovered.

15310 of 16716 relevant lines covered (91.59%)

3941593.23 hits per line