- Reliable document processing
- Only information available before the filing was used
- Evidence-linked change detection
- Historical comparison
- Reproducible classifications
Empirical journey
The market hypothesis narrowed. The workflow opportunity became clearer.
Three locked studies separated what the evidence supports from what remains unproven.
Negative and null results were preserved rather than tuned away.
Research in plain language
Capability is supported. Market and commercial claims are not.
This is the shortest accurate interpretation.
- Incremental news alpha
- Improved expected-language reconstruction
- Causal claims
- Commercial adoption or willingness to pay
- Validated downside-risk prediction
- Analyst time savings, trust, recall, and repeat use
Research decision tree
Modeling stopped when each evidence gate failed.
The null results did not end the project; they prevented unsupported claims.
Trading hypothesisDoes added document data improve prediction?
→Gate 1 · not supportedStop incremental-alpha claim
→Expected-language modelDoes extra data beat prior 10-K?
→Gate 2 · not supportedStop added model complexity
→Workflow productTest analyst value next
Study 1 · combined multisource model
Did 10-K information improve the 8-K and inexpensive-news model?
- Question
- Did adding 10-K features add one-month return information?
- What was tested
- Locked historical 8-K, 10-K, news, company-characteristic, and return panels with time-ordered folds.
- Result
- Adding the 10-K information did not improve the model once ordinary company characteristics were accounted for.
- Decision
- Not supported Do not advance an incremental-alpha claim.
- Lesson
- Attractive absolute statistics were not evidence that document information added value.
Technical metrics
Combined mean monthly Spearman IC: 0.0862. Incremental IC versus 8-K plus EODHD: −0.00227; one-sided p = 0.9531; two of five folds positive; 95% block interval [−0.00567, 0.00049]. Decision label: NOT SUPPORTED—EXPLORATORY HISTORICAL LOCKED.
Not claimed: alpha, causality, executability, live trading performance, or incremental combined-news information.
Study 2 · expected-disclosure model
Could other data predict the next 10-K better than the prior 10-K alone?
- Question
- Did company characteristics, intervening 8-Ks, and news improve the next filing’s representation?
- What was tested
- Consecutive filing representations and strictly prior characteristics, 8-Ks, and headlines.
- Result
- The prior 10-K alone predicted the next filing representation better than every augmented model.
- Decision
- Not supported Do not advance an enriched expected-language shock.
- Lesson
- Direct historical comparison was more defensible than added modeling complexity.
Method detail
All features obeyed point-in-time eligibility. The persistence benchmark won the locked comparison.
Not claimed: that news or characteristics add predictive information beyond the prior filing, or that representation error predicts returns.
Study 3 · disclosure-change demonstration
Can the system identify and evidence meaningful filing changes?
- Question
- Can consecutive 10-K differences be categorized, linked to exact evidence, and checked against earlier disclosures?
- What was tested
- 150 issuer-disjoint companies, 150 pairs, 995 retained changes, and eight outcome-independent cases.
- Result
- The evidence workflow operated reliably. The bounded risk comparison was inconclusive.
- Decision
- Supported product Advance to an analyst workflow pilot—not trading deployment.
- Lesson
- A product can reduce research friction even when a predictive market claim is not established.
Locked product counts
629 high-confidence changes; 1,990 verified excerpt offsets; 394 previously disclosed; 342 genuinely new relative to the bounded study sources; 243 partially anticipated; 16 unclear; zero timing violations. The novelty counts sum to 995.
Not claimed: proven alpha, causal or validated risk prediction, customer adoption, willingness to pay, or institutional validation.
Next evidence gate
Test the analyst workflow.
Measure speed, recall, correction burden, trust, repeat use, and stored-context value.