2b) SEC EDGAR datasets (Form 4 lines, 8-K events, XBRL fundamentals, Form D offerings)
Beyond forecast scorecards and classic 13F-style holdings maps, InsightMeter maintains authenticated research tables sourced from U.S. SEC public disclosure pipelines. These tables are education tools: they organize fields that already exist in primary filings and company-facts extracts. They are not a substitute for reading the accession on EDGAR when a claim becomes specific.
CIK↔ticker glue and EDGAR submission indexes help readers confirm issuer identity and open verify links to primary documents. Cached edgar_submissions coverage is currently on the order of ~300 CIKs (pilot expansion; approximate; subject to loader refresh). Company lookup also uses sec_company_tickers.
Form 3/4/5 line-level insider transactions expose transaction codes, amounts, and related fields. These line tables are loaded market-wide by quarter (not a tiny CIK pilot). The current loaded window spans filings about 2018-01-02 through 2026-09-30 with on the order of ~1,773,788 submissions (approximate; verified 2026-10-11 HKT; subject to loader refresh). Free accounts see a short educational sample; Pro unlocks longer lists and CSV export on that deeper window.
Form 8-K event indexes surface item codes with EDGAR links for covered issuers (currently on the order of ~591 CIKs—wider than the earlier ~187 pilot, still not a complete U.S. market). Loaded filing history currently reaches from about 1999 through recent 2026 filings (approximate; subject to loader refresh)—deeper history and wider issuer coverage, without claiming every ticker. Free sees a short sample; Pro unlocks longer lists and CSV.
XBRL fundamentals subset (for example Revenue, net income, EPS, assets, and related us-gaap tags) sits beside research for covered issuers (currently on the order of ~494 CIKs—wider than the earlier ~106 pilot, still not full-market). Company-facts end dates currently span about 2018 through Sep 2026 (approximate; subject to loader refresh). Free sees a latest-style snapshot sample; Pro unlocks longer history and CSV.
Form D exempt offerings join form_d_offering + form_d_issuer on accession (about ~472,993 offering rows as of the 2026-10-11 HKT snapshot). Fields include industry group, federal exemption items, total offering amount / amount sold / remaining, issuer entity name, state, entity type, and Verify on EDGAR links. Loaded filed dates currently span about 2020-07-01 through 2026-09-30 with source_quarter about 2020q3–2026q3; pre-2020q3 is empty for the current loader (approximate; subject to loader refresh). Free sees a short educational teaser; Pro/Elite unlock longer lists and CSV. Empty [] when a ticker or CIK has no Form D in the loaded quarters is normal—not a product failure and not a signal.
Daily stock_prices history currently spans about 2018-01-02 through 2026-10-09 on the order of ~9,678,121 rows (approximate; verified 2026-10-11 HKT). Derived insider_summary_new / related alpha summaries remain as-of 2026-06-07 because no newer summary CSVs have been loaded. Forecast predictions tables are unchanged in this refresh. Macro series (Treasury yields, NYFED SOFR, BLS CPI/unemployment/payrolls) ship as previously loaded.
Limitations are intentional and public: reporting lag, amendment risk, mapping errors, incomplete CIK width on 8-K/XBRL/edgar_submissions (Form 3/4/5 remains market-wide by quarter; Form D is quarter-loaded offering notices from ~2020q3 onward for the current loader, not a complete private-market census). Empty results outside a loaded set are normal. Approximate date windows move when loaders refresh. Soft Free/Pro sample lengths are unchanged. Soft product note: explore Free teasers to practice vocabulary; Pro (HK$40/mo) is optional for longer Form 4 / 8-K / fundamentals / Form D lists and CSV—never a guarantee of investment returns.
Roadmap language stays careful: coverage for 8-K and XBRL is wider than the early pilot and history is deeper—especially 8-K back toward ~1999—while Form 3/4/5 and Form D now sit on multi-year loaded windows (Form 3/4/5 from ~2018; Form D from ~2020q3 for the current loader). We still do not claim a complete U.S. market or a complete private-placement universe. Soft Free/Pro sample lengths are unchanged. Soft product note: explore the dashboard on Free; Pro (HK$40/mo) is optional when you need longer lists and CSV—never a guarantee of investment returns.
4b) Worked scoring example (fictional, education-first)
The following walkthrough uses fictional names and numbers to show how a forecast score can be computed when horizon and benchmark are explicit—and how the same numbers can mislead when survivorship or cherry-picked windows sneak in. It is an educational template, not a performance claim and not guaranteed alpha.
Setup (write these before you score)
Forecaster “River” publishes a directional call on fictional ticker MAPL. Locked definition: (1) Claim type = open price on event date T vs close on T+20 trading days; (2) Direction = “MAPL will outperform the fictional sector ETF SECT by more than 0% over that window” (excess-return hit, not absolute-up hit); (3) Hit = excess return > 0; miss = excess return ≤ 0; (4) Sample = all River MAPL-vs-SECT calls with the same definition in calendar years 2022–2024, including calls that expired worthless for the narrative. No discretionary dropping of quiet months.
Step-by-step hits and misses (fictional prints)
Call A (T=2024-01-10): MAPL +4.0% over 20 sessions; SECT +1.0%; excess +3.0% → hit. Call B (T=2024-03-05): MAPL −2.0%; SECT −5.0%; excess +3.0% → hit (absolute down, relative win—allowed only because the locked definition used excess return). Call C (T=2024-06-12): MAPL +6.0%; SECT +7.5%; excess −1.5% → miss. Call D (T=2024-09-20): MAPL +1.0%; SECT +1.0%; excess 0.0% → miss under a strict >0 hit rule.
Score for this four-call toy sample: 2 hits / 4 scored = 50% hit rate. Educational point: without the written definition, a marketer could relabel Call B a miss (“stock fell”) or Call C a hit (“stock rose”) and invent a prettier percentage. The methodology win is the lockfile, not the headline rate.
If InsightMeter-style summaries show accuracy metrics, they should remain tied to fixed windows and restated when corrections arrive. Users should be able to ask: what horizon, what benchmark, what hit rule, what sample?
Failure modes to keep visible
Survivorship: dropping Call C and Call D because they “weren’t real setups” after seeing outcomes. The cleaned deck jumps to 100% and teaches nothing reliable.
Cherry-picked windows: keeping the same four calls but rescoring with a 5-day horizon only on the days that help, or switching from SECT excess to absolute return after browsing results.
Silent benchmark swaps: advertising a sector-excess track record while implying absolute “stock went up” skill in the next pitch.
Tiny samples: treating 2/4 as a stable edge. Education favors showing sample size beside any rate.
Look-ahead and recycled screenshots: scoring with prices that were not knowable under the stated timestamp rules.
Use this example as a checklist when you read any forecast scoreboard—ours or elsewhere. Past hit rates are historical descriptions under a definition. They are not promises of future performance, trading advice, or guaranteed alpha.