ML Roadmap — Insider Intel

The constraint

The bottleneck is time, not engineering

Learning from outcomes requires outcomes to exist. Price returns at t+30, t+60, and t+90 days take 30 to 90 days to materialise after a signal is issued. Data collection started now; the first meaningful regression runs in 3–4 months. There's no shortcut — and any system claiming instant weight adaptation is working on look-ahead data or noise.

✓

Phase 01

Outcome Tracking

● Live now

Since August 2026 · accumulating daily

Every AI Signal scored today is written to a SignalOutcome record — all six factor scores, the composite, the signal label, and the price at the time of issue. A daily scheduler then checks open records and fills in actual price returns at t+30, t+60, and t+90 days automatically. This is the ground-truth dataset everything else depends on.

SignalOutcome table: ticker, signal date, all 6 factor scores, composite, signal label

Price at signal date + returns filled automatically at 30 / 60 / 90 days

S&P 500 benchmark tracking — SPY price recorded at signal time, returns at t+30/60/90, beat_30d/60d/90d booleans

Idempotent daily scheduler — no duplicate records, no gaps

GET /api/signals/outcomes endpoint for internal review

Phase 02

Factor Correlation Analysis

○ ~6 weeks

First 30-day returns available mid-September 2026

Once return_30d values begin populating, run Pearson correlation between each factor score and actual returns. This answers the most important question before any optimisation: do the current weights correspond to what actually predicts returns? No model yet — pure statistics, the first empirical sanity check.

Weekly scheduled job computing per-factor Pearson r against return_30d, return_60d, return_90d

factor_correlations table — one row per factor per week

Internal analytics view: factor correlation trends over time

Factor Correlation Dashboard

Populates ~Sep 2026

FactorWeightr (30d)r (60d)r (90d)Signal

Insider Buy Activity

25%

—

No data

Analyst Consensus

25%

—

No data

Price Target Upside

20%

—

No data

AI Sentiment

20%

—

No data

Insider Track Record

15%

—

No data

Buying Into Weakness

10%

—

No data

Correlation values will populate once 10+ signals have 30-day returns — expected mid-September 2026.

Phase 03

Weight Optimisation via Regression

○ ~Nov 2026

Pipeline built · waiting for ~300 matured signal records (~Nov 2026)

Use accumulated SignalOutcome rows as training data. Features: the six factor scores. Target: return_30d (or a binary — did the signal beat the S&P 500?). Fit a constrained linear regression (weights must be positive and sum to 1). The resulting coefficients replace the hardcoded weights in the scoring engine. This is accurate to call self-improving.

Pure Python projected gradient descent onto the probability simplex — no external ML dependencies

Constrained optimisation: weights > 0, sum to 1 (simplex projection algorithm)

Walk-forward validation — oldest 80% train, newest 20% validate, no look-ahead bias

Synthetic test mode — 300 samples with known true weights to validate the pipeline now

factor_weights table — every run recorded with date, sample size, train/val R², and is_live flag

Naming note: This is statistical weight learning — constrained gradient descent on a linear model, not a neural network. "Self-improving" is accurate here. "Backpropagation" is not — that term requires a neural net (see Phase 4).

Phase 04

Neural Net & Backpropagation

○ ~6–9 months

Justified only once regression proves insufficient — needs 500+ diverse samples

Linear regression assumes factor contributions are additive and independent. They may not be — a high insider_score may only matter when track_record_score is also high. A small MLP (multilayer perceptron) captures these interaction effects. 6 inputs → 8 hidden units → 1 predicted return. Trained via gradient descent and backpropagation. This is the phase where "backpropagation" is accurate to claim.

PyTorch or scikit-learn MLPRegressor — small architecture (6 → 8 → 1)

Training pipeline with walk-forward cross-validation

Comparison against Phase 3 regression — only deployed if MLP outperforms

Model weights versioned and stored; rollback available at any time

SHAP values for interpretability — factor contributions still visible per signal

Gate: Phase 4 only runs if Phase 3 regression plateaus or underperforms a naive benchmark. A neural net that doesn't outperform linear regression isn't worth the added complexity — and adds explainability cost.

Current data accumulation

~20 / day

Signal outcomes recorded (top-20 tickers from rolling 30-day window)

30 days

Until first return_30d values are available for correlation analysis

300–500

Signal records needed before regression weights are statistically meaningful

500+

Diverse samples across market conditions before a neural net is justified

Product Roadmap

What we're building next

We're sharing this publicly because we believe in building with our users, not just for them. This is a living view of our priorities — shaped by the feedback we receive from early access members.

Recently Shipped

Shipped

Alert Rules Engine

Configurable alert rules across five trigger types — new SEC filings, open-market buys, price movements, cluster buys, and insider watchlist hits. Each rule supports threshold filters, role filters, and per-rule delivery preferences.

AlertsRulesTriggers

Shipped

Email & Slack Alert Delivery

Alert notifications delivered via email and Slack webhook — filed activity reaches your inbox or channel within minutes of SEC publication. Configurable per-organisation with a daily volume cap to protect deliverability during the alpha period.

AlertsEmailSlack

Shipped

Real-time In-App Notifications & Alert Log

Instant WebSocket-delivered toast notifications when an alert fires, with a persistent alert log and unread badge count that survives page reloads. No polling — alerts appear the moment they are triggered.

AlertsReal-timeWebSocket

Shipped

Insider Track Record Scoring

Each insider's historical open-market buy accuracy is scored against subsequent 30-, 60-, and 90-day price performance and incorporated as a weighted factor in the AI Signal composite. See whether a CFO's past buys have preceded meaningful moves — and let that history influence the signal strength.

InsidersTrack RecordAI Signal

Shipped

Team Accounts & Member Invitations

Invite colleagues to a shared workspace via email. Role-based access (owner, admin, member) with seat limits enforced per plan tier. Org admins manage delivery settings and alert rules centrally for the whole team.

TeamsCollaborationRoles

Up Next

Browser Push Notifications

Native browser push notifications so alert fires reach you even when the app isn't open in your active tab — no email round-trip required. Opt-in per device, with the same per-rule delivery controls as email and Slack.

AlertsPushBrowser

Up Next

Group Consensus Tracker

Create a group, invite friends or colleagues, and track your collective view on insider activity alongside your own. Each member votes Bullish, Neutral, or Bearish on a ticker — the platform surfaces the group consensus, flags divergence, and lets you add a note to your signal. Built for investment clubs, family offices, and small funds who want a shared lens on the same data.

GroupsCollaborationSignals

Planned

Shared Alert Rules & Watchlists

Team plan accounts will be able to create alert rules and watchlists that are shared across all members of the workspace — so the whole team fires on the same signals without each member duplicating configuration. Included in the Team plan tier.

TeamsAlertsWatchlists

Planned

Insider Cluster Detection

Automatically surface tickers where multiple insiders transacted within a short window — a historically strong signal of informed conviction. The AI Insights layer will annotate clusters with context about the insiders involved and their historical accuracy.

AI InsightsPattern DetectionDashboard

Planned

Portfolio Overlap Analysis

Map your watchlist against your existing portfolio holdings and highlight where insider activity overlaps with positions you already carry. Designed for RIAs managing multiple client accounts who need a fast view of material non-public activity risk.

PortfolioRiskRIA

Planned

Data-as-a-Service API

Programmatic access to the full Insider Intel dataset — real-time Form 4 events, AI-generated summaries, and insider track records via REST and webhook. Aimed at quant teams, fintech builders, and research firms who want to integrate insider signals into their own workflows.

APIWebhooksEnterprise

Exploring

13F Institutional Overlap

Cross-reference insider buying with 13F filings to reveal where institutional money and insider conviction align — one of the strongest combined signals in fundamental analysis.

13FInstitutionalSignal

Exploring

Earnings Calendar Integration

Flag insider transactions that occur in the weeks before an earnings date — a key context signal. Highlight trades that fall inside or outside the typical pre-earnings quiet period for each company.

EarningsCalendarContext

Exploring

Mobile App

A native iOS and Android companion for monitoring your watchlist and receiving real-time alerts on the go. The web platform will always be the primary surface; the mobile app is a companion, not a replacement.

iOSAndroidAlerts

Exploring

International Market Expansion

The SEC Form 4 disclosure regime has equivalents in every major market — each with its own filing format, regulator, and timing rules. We're exploring coverage of three regions as the first expansion beyond US equities.

🇦🇺 ASX — Australia Appendix 3Y

Directors and officers of ASX-listed companies must lodge an Appendix 3Y with the ASX within 5 business days of any change in their relevant interests. Disclosed through the ASX Market Announcements Platform — publicly accessible, machine-readable, and structurally close to Form 4.

🇬🇧 FCA — United Kingdom PDMR Notifications

Under UK MAR (Market Abuse Regulation), Persons Discharging Managerial Responsibilities (PDMRs) and their closely associated persons must notify both the company and the FCA within 3 business days of transactions exceeding €5,000. Disclosures are published via the Regulatory News Service (RNS) and the FCA's National Storage Mechanism.

🇪🇺 ESMA — European Union Directors' Dealings

EU MAR imposes the same 3-business-day PDMR disclosure obligation across all 27 member states, with filings published through each national competent authority (e.g. AMF in France, BaFin in Germany, ESMA's central register). Coverage here means parsing disclosures across multiple jurisdictions and languages — technically feasible, operationally complex.

Priority order, data accessibility, and which exchanges to cover first will be shaped by user demand. If international coverage matters to you, tell us which market — it moves the timeline.

ASXUK MAREU MARPDMRGeo-expansion

Shape what we build

Have a feature request, a workflow we haven't covered, or feedback on what's already live? We read every message.

From fixed weights to learned ones