PXI /COMMAND

A suite of quantitative market tools:

• PXI Index - Real-time macro market strength (28 indicators → single 0-100 score)
• Sector Signals - Weekly sector rotation signals from Reddit discussion analysis

Live: pxicommand.com · /signals

What is PXI?

PXI (Positioning Index) synthesizes signals from liquidity, credit spreads, volatility, market breadth, macro economic data, global risk appetite, and crypto markets into one score:

Score	Status	Meaning
80+	MAX PAMP	Extremely bullish conditions
65-79	PAMPING	Bullish conditions
45-64	NEUTRAL	Mixed signals
30-44	SOFT	Bearish conditions
<30	DUMPING	Extremely bearish conditions

Features

Core Index

• 28 indicators across 7 categories normalized to 0-100 scale
• Real-time updates via GitHub Actions cron (4x daily)
• Historical backtest with 1000+ observations since Dec 2022

ML & Predictions (v1.2)

• Ensemble Predictions - Weighted combination of XGBoost (60%) and LSTM (40%) with confidence scoring
• XGBoost Model - Gradient boosted trees using 36 engineered features (momentum, dispersion, extremes)
• LSTM Model - Recurrent neural network using 20-day sequences of 12 features for temporal patterns
• Similar Period Detection - Find historically similar market regimes using vector embeddings
• Prediction Tracking - Log predictions and evaluate against actual outcomes
• Historical Outlook - Forward returns and win rates by PXI bucket with adaptive thresholds
• Signal Layer - Risk allocation signals based on PXI, momentum, and volatility

Backtesting

• PXI-Signal Strategy - Dynamic risk allocation (Sharpe ~2.0)
• Comparison - vs 200DMA and buy-and-hold baselines
• Walk-forward validation - Out-of-sample testing

MCP Server (AI Agent Integration)

• Model Context Protocol - Enables Claude and other LLM agents to query PXI
• 7 Tools - get_pxi, get_predictions, get_similar_periods, get_signal, get_regime, get_market_context, get_history
• Agent-Optimized - Structured responses with suggested actions and risk assessments

Sector Rotation Signals (/signals)

• Weekly Analysis - Runs every Monday (or Tuesday if Monday is a US market holiday)
• Reddit Sentiment - Analyzes 750+ posts from investing subreddits
• 20 Sector Themes - Nuclear, Automation, Defense, Copper, Financials, and more
• Classification System - Signal type (Rotation/Momentum/Divergence/Reversion), confidence, timing
• Evidence-Based - Links to source discussions for each theme

Categories & Weights

Category	Weight	Key Indicators
Credit	20%	HY spreads, IG spreads, yield curve
Liquidity	15%	Fed balance sheet, TGA, reverse repo, net liquidity
Volatility	15%	VIX, VIX term structure
Breadth	15%	RSP/SPY ratio, advance-decline
Positioning	15%	Put/call ratio, fear & greed
Macro	10%	10Y yield, dollar index, oil
Global	10%	Copper/gold ratio, BTC, stablecoin flows

Architecture

                              PXI INDEX                                    SIGNALS
                    ┌─────────────────────────┐                 ┌─────────────────────────┐
                    │                         │                 │                         │
┌─────────────────┐ │  ┌──────────────────┐   │  ┌───────────┐  │  ┌──────────────────┐   │
│   Data Sources  │───▶│  GitHub Actions  │   │  │  Reddit   │────▶│  Cloudflare Cron │   │
│  FRED, Yahoo,   │ │  │   (Daily Cron)   │   │  │   API     │  │  │   (Weekly)       │   │
│  DeFiLlama...   │ │  └────────┬─────────┘   │  └───────────┘  │  └────────┬─────────┘   │
└─────────────────┘ │           │             │                 │           │             │
                    │           ▼             │                 │           ▼             │
                    │  ┌──────────────────┐   │                 │  ┌──────────────────┐   │
                    │  │  Cloudflare D1   │   │                 │  │  Cloudflare D1   │   │
                    │  │   (SQLite)       │   │                 │  │  + R2 Storage    │   │
                    │  └────────┬─────────┘   │                 │  └────────┬─────────┘   │
                    │           │             │                 │           │             │
                    └───────────┼─────────────┘                 └───────────┼─────────────┘
                                │                                           │
                                ▼                                           ▼
                    ┌──────────────────────────────────────────────────────────────────┐
                    │                     React Frontend (CF Pages)                     │
                    │                        pxicommand.com                             │
                    └──────────────────────────────────────────────────────────────────┘

Project Structure

pxi-command/
├── frontend/        # React app (pxicommand.com)
├── worker/          # Main PXI API worker
├── signals/         # Sector rotation signals worker (/signals)
├── ml/              # ML model training (XGBoost, LSTM)
├── mcp-server/      # MCP server for AI agents
├── src/             # Shared data collection scripts
└── ai/              # AI analysis prompts

Tech Stack

• Backend: TypeScript, Node.js
• Database: Cloudflare D1 (SQLite at edge)
• Storage: Cloudflare R2 (report storage for signals)
• Vector Store: Cloudflare Vectorize (768-dim embeddings)
• ML Models: XGBoost + LSTM (trained locally, inference at edge via KV)
• AI: Cloudflare Workers AI (BGE embeddings, Llama analysis)
• API: Cloudflare Workers
• Frontend: React 19, Vite, Tailwind CSS
• Hosting: Cloudflare Pages
• Scheduler: GitHub Actions (PXI daily) + Cloudflare Cron (Signals weekly)
• Agent Integration: MCP Server (Model Context Protocol)

API Endpoints

Core Data

Endpoint	Method	Description
/api/pxi	GET	Current PXI score, categories, sparkline, regime, and freshness operator payload (topOffenders, refresh ETA)
/api/history	GET	Historical PXI scores
/api/regime	GET	Current market regime detection
/api/signal	GET	PXI-Signal layer with risk allocation
/api/plan	GET	Canonical decision object for homepage (policy state, uncertainty, consistency, trader playbook)
/api/market/consistency	GET	Latest decision consistency score/state/violations
/api/ops/freshness-slo	GET	Rolling 7d/30d freshness SLO attainment and recent critical stale incidents
/api/ops/utility-funnel	GET	Rolling utility funnel metrics (session -> decision views -> no-action unlock coverage + CTA intent rate)
/api/ops/decision-impact	GET	Decision-impact ops view (7d/30d market outcome proxy, theme summary, utility attribution, governance mode with observe + enforce readiness)
/api/ops/decision-grade	GET	Rolling governance scorecard (freshness, consistency, calibration, edge evidence, opportunity hygiene, utility)
/api/ops/go-live-readiness	GET	Explicit go-live readiness snapshot (blockers + decision-impact enforce readiness context)

Product Layer (Phase 1)

Endpoint	Method	Description
/api/brief	GET	Daily market brief coherent with /api/plan policy state (includes contract version + consistency)
/api/opportunities	GET	Ranked opportunities for 7d or 30d horizon with calibration + expectancy + unavailable reasons
/api/decision-impact	GET	Outcome attribution using matured opportunity ledger rows (market uses SPY forward proxy; theme uses proxy blend with SPY fallback, 30/90d windows)
/api/diagnostics/calibration	GET	Calibration diagnostics (Brier/ECE/log loss + quality band) for conviction and edge-quality snapshots
/api/diagnostics/edge	GET	Forward-chaining edge evidence (model vs lagged baseline uplift, CI, leakage sentinel, promotion gate)
/api/alerts/feed	GET	In-app alert timeline (regime_change, threshold_cross, opportunity_spike, freshness_warning)
/api/alerts/subscribe/start	POST	Start email digest subscription with verification token
/api/alerts/subscribe/verify	POST	Verify subscription token and activate email digest
/api/alerts/unsubscribe	POST	Unsubscribe via token
/api/metrics/utility-event	POST	Record lightweight product utility telemetry events (session, decision view, no-action unlock, cta_action_click)

ML & Predictions

Endpoint	Method	Description
/api/similar	GET	Find similar historical periods (vector search with weighted scoring)
/api/predict	GET	Historical outlook by PXI bucket with adaptive thresholds
/api/ml/predict	GET	XGBoost model predictions for 7d/30d PXI changes
/api/ml/lstm	GET	LSTM neural network predictions using 20-day sequences
/api/ml/ensemble	GET	Weighted ensemble (60% XGBoost + 40% LSTM) with confidence
/api/accuracy	GET	Prediction accuracy metrics
/api/analyze	GET	AI-generated market analysis

Backtesting

Endpoint	Method	Description
/api/backtest	GET	PXI bucket forward returns analysis
/api/backtest/signal	GET	Signal strategy vs baselines

Admin (requires auth)

Endpoint	Method	Description
/api/write	POST	Write indicator data
/api/recalculate	POST	Recalculate PXI for date
/api/migrate	POST	Create missing database tables
/api/evaluate	POST	Evaluate past predictions
/api/retrain	POST	Retrain adaptive bucket thresholds
/api/export/training-data	GET	Export data for ML model training
/api/market/refresh-products	POST	Generate/store latest brief, opportunities, and market alerts
/api/market/send-digest	POST	Send daily digest to active subscribers

Signals (/signals/*)

Endpoint	Method	Description
/signals/latest	GET	Latest weekly report (HTML)
/signals/run/{id}	GET	Specific report by run ID
/signals/api/runs	GET	List all runs with metadata
/signals/api/runs/{id}	GET	Run details as JSON
/signals/api/run	POST	Trigger manual run (requires X-Admin-Token)
/signals/og-image.png	GET	Dynamic OG image for social sharing

Agent Docs

Agent-discovery and markdown docs are published as static frontend endpoints:

• https://pxicommand.com/llms.txt
• https://pxicommand.com/agent.md
• https://pxicommand.com/signals/agent.md

Canonical read endpoints referenced by llms.txt:

• https://api.pxicommand.com/health
• https://api.pxicommand.com/api/pxi
• https://pxicommand.com/signals/api/runs
• https://pxicommand.com/signals/api/accuracy
• https://pxicommand.com/signals/api/predictions

Write/admin endpoints remain auth-protected and are not intended for autonomous agent writes.

Data Freshness SLA

npm run cron:daily enforces per-indicator freshness SLAs before any write/recalculate calls.

SLA Classes

Class	Max age (days)
daily	4
realtime	4
weekly	10
monthly	45
source-lagged (wti_crude)	7
source-lagged (dollar_index)	10

Explicit Threshold Overrides

Indicator	Max age (days)
cfnai	120
m2_yoy	120
fed_balance_sheet	14
treasury_general_account	14
reverse_repo	14
net_liquidity	14

Critical SLA Gate

The daily refresh fails if any of these are stale or missing after fetch stage:

• aaii_sentiment
• copper_gold_ratio
• vix
• spy_close
• dxy
• hyg
• lqd
• fear_greed

Run summary output is written to /tmp/pxi-sla-summary.json and published to GitHub Actions step summary.

Market Product Automation

• Daily PXI Refresh now runs ingestion + SLA gate + post-refresh product generation (/api/market/refresh-products).
• Product summary output is written to /tmp/pxi-market-summary.json and appended to the Actions job summary.
• A separate Daily PXI Digest workflow runs at 13:00 UTC (8:00 AM ET standard) and calls /api/market/send-digest.
• Digest summary output is written to /tmp/pxi-digest-summary.json.

PXI Audit Findings Remediation

Taylor’s PXI audit findings are being remediated with trust-first controls:

1. Opportunity coherence gate

• Directional opportunities are only published when contract checks are coherent.
• Bullish items must have non-negative expectancy and p_correct >= 0.50 when available.
• Bearish items must have non-positive expectancy and p_correct >= 0.50 when available.
• Neutral items are not published in phase 1 actionable feed.

2. Degraded-mode suppression

• /api/opportunities is hard-suppressed (items: []) when:
  • critical_stale_count > 0, or
  • market consistency state is FAIL.
• Response includes suppressed_count and degraded_reason (for example suppressed_data_quality, coherence_gate_failed, quality_filtered).

3. Calibration diagnostics transparency

• New endpoint: GET /api/diagnostics/calibration
• Includes Brier score, ECE, log loss, quality band, and insufficient-sample reasons.

4. Freshness reliability governance

• New endpoint: GET /api/ops/freshness-slo
• Reports rolling 7d and 30d SLO attainment using market_refresh_runs, including recent critical stale incidents.

5. Edge-proof diagnostics + promotion gate

• New endpoint: GET /api/diagnostics/edge
• Reports forward-chaining directional uplift vs lagged baseline, CI bounds, and leakage sentinel health per horizon.
• /api/market/refresh-products now fails closed when leakage sentinel checks fail, blocking publish/promotion.

6. Decision-stack unification + cross-horizon coherence

• /api/plan now includes additive decision_stack and cross_horizon blocks.
• cross_horizon.state (ALIGNED|MIXED|CONFLICT|INSUFFICIENT) downgrades ACTIONABLE -> WATCH on conflict/insufficient evidence.
• Additional invalidation guidance is attached when horizons conflict or one horizon is missing.

7. Signals sanitation + stability

• Adapter-side ticker sanitation (allowlist regex, jargon stopwords, dedupe/cap) before PXI opportunity blending.
• Signals source velocity cap tightened (GROWTH_RATIO_CAP = 25) with explicit growth_ratio_capped flag in metrics payload.

8. Opportunity TTL timebox + cache safety

• /api/opportunities now emits refresh recency metadata:
  • data_age_seconds
  • ttl_state (fresh|stale|overdue|unknown)
  • next_expected_refresh_at
  • overdue_seconds
• CTA is hard-disabled when ttl_state is overdue or unknown.
• Cache policy is timebox-aware: overdue/unknown states return Cache-Control: no-store to avoid stale hard-CTA exposure.

9. Utility funnel instrumentation (Phase 3)

• New ingestion endpoint: POST /api/metrics/utility-event.
• New ops endpoint: GET /api/ops/utility-funnel?window=7|30.
• Frontend now emits additive utility events for:
  • session_start
  • plan_view
  • opportunities_view
  • decision-state exposures (decision_actionable_view|decision_watch_view|decision_no_action_view)
  • no_action_unlock_view (tracks no-action days as positive threshold-communication workflow events)
  • cta_action_click (anonymous action intent from /opportunities)

10. Decision-grade governance (Phase 4)

• Immutable opportunity ledger tracks candidate-vs-published counts per horizon on every product refresh.
• New ops endpoint: GET /api/ops/decision-grade?window=7|30.
• Scorecard components:
  • freshness SLO reliability
  • consistency-state quality
  • calibration quality band health
  • edge-promotion evidence
  • opportunity hygiene (over-suppression + cross-horizon conflict persistence)
  • utility funnel adoption

11. Outcome attribution + utility lift (Phase 5)

• New endpoint: GET /api/decision-impact?horizon=7d|30d&scope=market|theme&window=30|90
• New endpoint: GET /api/ops/decision-impact?window=30|90
• Refresh pipeline persists:
  • item-level publish/suppression ledger (market_opportunity_item_ledger)
  • decision-impact snapshots (market_decision_impact_snapshots)
• Attribution basis is spy_forward_proxy (market and theme scope in this phase).
• Governance is observe-only in this phase. Breaches are surfaced in ops/workflow summaries but do not fail refresh runs.
• Observe thresholds:
  • market_7d.hit_rate >= 0.52
  • market_30d.hit_rate >= 0.50
  • market_7d.avg_signed_return_pct > 0
  • market_30d.avg_signed_return_pct > 0
  • cta_action_rate_pct >= 2.0

12. Decision-impact enforcement + attribution hardening (Phase 6)

• ops/decision-impact now reports governance mode (observe|enforce), enforce readiness, and enforce breach list/count.
• Enforce mode is sample-gated and disabled by default:
  • FEATURE_ENABLE_DECISION_IMPACT_ENFORCE=false
  • DECISION_IMPACT_ENFORCE_MIN_SAMPLE=30
  • DECISION_IMPACT_ENFORCE_MIN_ACTIONABLE_SESSIONS=10
• Theme attribution basis upgraded from pure SPY proxy to theme proxy blend (credit/vol/global/crypto mappings) with explicit SPY fallback when proxy coverage is unavailable.
• Utility funnel CTA denominator now uses actionable sessions including CTA-click sessions (not only explicit actionable-view events) to reduce denominator undercount.
• POST /api/market/backfill-products can now rebuild opportunity ledgers and regenerate decision-impact snapshots from historical opportunity snapshots (rebuild_ledgers, default true).

13. Go-live readiness + maturity-window correction (Phase 7)

• Decision-impact sampling now uses maturity-window semantics (outcome date in-window) instead of raw as_of filtering, so 30d attribution no longer collapses to zero in a 30d window.
• /api/ops/decision-grade now emits explicit go_live_blockers[] plus readiness context (decision_impact_enforce_ready, sample counts, minimum thresholds).
• New endpoint: GET /api/ops/go-live-readiness?window=30 for a compact operator view of readiness blockers and score context.

14. Go-live blocker reduction + enforce readiness normalization (Phase 8)

• Utility decision target now scales by observed days in-window (instead of fixed full-window counts), improving early-window status stability.
• Decision-impact actionable-session minimum is normalized by observed days (with configured minimum still reported) to avoid false readiness negatives during partial windows.
• Go-live blocker logic now keys edge on promotion-gate safety (edge_promotion_gate_fail) rather than strict edge component pass/watch state.
• Enforcement should be flipped only when /api/ops/go-live-readiness shows go_live_ready=true and no blockers.

Scheduler Ownership Runbook

Target ownership is GitHub Actions (Daily PXI Refresh) only. Keep Cloudflare cron enabled during hardening, then cut over after stability validation.

72-hour Shadow Procedure

1. Keep Cloudflare and GitHub schedulers active.
2. Require 12 consecutive successful scheduled GitHub runs (4/day over 72 hours) with zero critical SLA violations.
3. Disable Cloudflare cron triggers for:
   • pxi-api-production
   • pxi-api
4. Re-verify Cloudflare schedules are empty and GitHub continues refreshing data.

Development

# Install dependencies
npm install

# Set up environment
cp .env.example .env
# Edit .env with your credentials

# Run daily data refresh
npm run cron:daily

# Start frontend dev server
cd frontend && npm run dev

# Deploy main worker
cd worker && npx wrangler deploy --env production

# Deploy signals worker
cd signals && npm install && npx wrangler deploy --env production

Launch Readiness Smoke Checks

Use these checks after deployment or before releases:

# Frontend route checks
curl -I https://pxicommand.com/
curl -I https://pxicommand.com/spec
curl -I https://pxicommand.com/alerts
curl -I https://pxicommand.com/guide
curl -I https://pxicommand.com/signals
curl -I https://pxicommand.com/signals/latest

# API checks (canonical host)
curl -I https://api.pxicommand.com/health
curl -I https://api.pxicommand.com/api/pxi
curl -I https://api.pxicommand.com/api/alerts
curl -I https://api.pxicommand.com/api/signal
curl -I https://api.pxicommand.com/api/diagnostics/edge?horizon=all

# CORS preflight check
curl -X OPTIONS https://api.pxicommand.com/api/refresh \
  -H "Origin: https://pxicommand.com" \
  -H "Access-Control-Request-Method: POST" \
  -H "Access-Control-Request-Headers: Content-Type, Authorization, X-Admin-Token"

# Backfill validation check (expect 400/401 depending on auth mode)
curl -X POST https://api.pxicommand.com/api/backfill \
  -H "Content-Type: application/json" \
  -d '{"start":"invalid-date","limit":"bad"}'

Temporary rollback path remains available via https://pxi-api.novoamorx1.workers.dev during cutover.

Signals Worker Route Check

Confirm the route is attached and active:

cd /Users/scott/pxi/signals
npx wrangler whoami
npx wrangler deployments list --name pxi-signals --env production

If /signals is unreachable (404), verify:

1. signals/wrangler.toml has pxicommand.com/signals and pxicommand.com/signals/* route entries.
2. The worker deploy succeeds after route changes.
3. The pxi-signals worker has active production bindings (d1, r2, kv) for pxicommand.com.

Signals Development

cd signals

# Run tests (258 tests)
npm test

# Generate offline report (uses sample data)
npm run offline

# Deploy to Cloudflare
npx wrangler deploy --env production

# Set admin token for manual runs
npx wrangler secret put ADMIN_RUN_TOKEN

ML Model Training

# Install Python dependencies
cd ml && pip install -r requirements.txt

# Train XGBoost models (requires WRITE_API_KEY for data export)
export WRITE_API_KEY=your_key
python train_xgboost.py

# Train LSTM models (requires PyTorch)
pip install torch
python train_lstm.py

# Upload models to Cloudflare KV
npx wrangler kv key put "pxi_model" --path=pxi_model_compact.json \
  --namespace-id=88901ff0216a484eb81b8004be0f5aea --remote
npx wrangler kv key put "pxi_lstm_model" --path=pxi_lstm_compact.json \
  --namespace-id=88901ff0216a484eb81b8004be0f5aea --remote

XGBoost (train_xgboost.py):

• Engineers 36 features (momentum, dispersion, extremes, rolling stats)
• Gradient boosted trees for 7d/30d PXI change prediction
• Exports to JSON (318KB compact)

LSTM (train_lstm.py):

• Uses 20-day sequences of 12 features (PXI, categories, VIX, dispersion)
• Single-layer LSTM with 32 hidden units
• Exports weights to JSON for edge inference (239KB compact)

MCP Server Setup

# Build the MCP server
cd mcp-server && npm install && npm run build

Claude Desktop - add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "pxi": {
      "command": "node",
      "args": ["/path/to/pxi/mcp-server/dist/index.js"]
    }
  }
}

Available Tools:

Tool	Description
get_pxi	Current score, categories, trend
get_predictions	ML ensemble forecasts (7d/30d)
get_similar_periods	Vector similarity search
get_signal	Risk allocation signal
get_regime	Market regime analysis
get_market_context	Comprehensive agent context
get_history	Historical scores

Environment Variables

Variable	Description
FRED_API_KEY	FRED API key (get one free)
WRITE_API_KEY	API key for write operations
WRITE_API_URL	Worker API URL for cron

Cloudflare Secrets (set via wrangler)

npx wrangler secret put WRITE_API_KEY
npx wrangler secret put FRED_API_KEY

Database Schema

Core Tables

• indicator_values - Raw indicator data from sources
• indicator_scores - Normalized 0-100 scores
• category_scores - Category-level aggregations
• pxi_scores - Final composite scores

ML Tables

• prediction_log - Predictions vs actual outcomes
• model_params - Tunable model parameters
• period_accuracy - Historical period prediction quality
• market_embeddings - Vector embeddings for similarity search

Signal Tables

• pxi_signal - Risk allocation signals
• alert_history - Generated alerts
• backtest_results - Strategy performance

Data Sources

• FRED - Federal Reserve economic data (liquidity, credit spreads, yields)
• Yahoo Finance - Market prices, ETFs, VIX, currencies
• DeFiLlama - Stablecoin market cap
• Coinglass - BTC funding rates
• CNN - Fear & Greed index

License

MIT