The project combines a completed financial-news study with the benchmark and research tools developed along the way. These examples explain how I handled data quality, reproducibility and financial accounting. The linked tests document the corresponding contracts; they do not establish predictive performance.
Problem. The legacy Kaggle merge contained 514 negative PhraseBank sentences a second time with incorrect neutral labels. Treating these conflicts as annotation ambiguity would have produced a misleading comparison. Excluding all conflicting groups would also have removed every PhraseBank negative from that comparison.
Implementation. I retained the original file for audit and rebuilt a derived benchmark from Financial PhraseBank and FiQA sources. The build preserves source identity, PhraseBank agreement tiers and FiQA continuous scores alongside the classification labels. The default dataset contains 5,947 unique sentences with zero conflicting duplicate groups.
Verification. The validator reports class counts, duplicate groups and label conflicts. Dataset tests check both the clean default and the legacy file, keeping the known corruption visible. This checks provenance and consistency; it does not make every individual sentiment label unambiguous.
Build script · Dataset card · Tests
Problem. Results are hard to compare when model identity, prompts, input data or configuration change without being recorded. Resuming a partially completed pipeline can also mix incompatible artifacts.
Implementation. The benchmark records prompt hashes and runtime metadata. The separate historical strategy pipeline derives run identities from its configuration and input identities, and records stage manifests with output hashes. Completed stages validate existing artifacts before reuse and reject mismatches. Atomic writes protect readers from partially written files.
Verification. Tests exercise interrupted-stage resumption, matching-stage reuse, changed-input rejection, output tampering and concurrent immutable writes. The offline demo uses fresh output directories so it cannot reuse a historical run. Hashes establish artifact identity, not scientific validity.
Stage manifests · Atomic I/O · Tests
Problem. Summing event returns can obscure how much capital is committed when holding periods overlap. A strategy can look attractive before transaction costs while failing after realistic accounting.
Implementation. The funded portfolio evaluator allocates overlapping cohorts through rotating capital sleeves, marks positions daily, applies costs and reconciles stock-level profit and loss to portfolio net asset value. Concentration limits apply across overlapping sleeves; unused capacity remains in cash.
Verification. Tests reconcile holdings, daily marks, terminal profit and loss, costs and net asset value. They also check concentration limits and the strict exchange-open availability boundary. The dissertation separately reports missing row-level news availability times as a binding limitation on predictive claims.
Portfolio evaluator · Tests · Final research design
The training-period association did not replicate in the later evaluation. Timing diagnostics and costs further limited the interpretation. The completed project therefore demonstrates a documented evaluation process with null and adverse findings preserved. The synthetic demo illustrates software execution only.