marketXED is a decision-support tool for short-term US equity trades. It continuously scans a watchlist plus dynamic Yahoo screeners, scores each ticker through a committee of independent analytical agents, and produces a small number of actionable signals per session, each accompanied by an explicit, beginner-safe playbook (limit price, stop, target, share count, dollar risk, exit time).

• It is a structured second opinion that enforces process: defined risk, hard exits, calibrated probabilities, and a learning feedback loop.
• It is not a broker, an execution engine, an oracle, or a guarantee. You place every order. You own every loss and every gain.

Every poll cycle (~60 seconds) the server performs roughly the following work for each ticker in the active universe:

1. OHLCV candles are fetched from Yahoo Finance (free, public). Both intraday and daily series are pulled depending on mode (Day vs Swing).
2. Sentiment is computed from X/Twitter posts (when feed is enabled) using VADER, an off-the-shelf NLP scorer tuned for short social text. Score range: −1 (very negative) to +1 (very positive).
3. Regime detection classifies the broader market (trending / mean-reverting / volatile / quiet) and scales each agent's vote accordingly.
4. Vetoes filter out structurally unsuitable setups (e.g. ultra-thin volume, illiquid pre-market, gap-and-fade traps).
5. Learning weights from the historical hit-rate of each agent on each regime are applied as multipliers.

All persistent files (state.json, signals.json, learning.json, predictions.jsonl) are written atomically with secrets locked to 0o600 permissions.

Ten independent agents each emit a vote in [-1, +1] and a confidence in [0, 1]. They are deliberately heterogeneous so they fail in different ways:

The committee then runs a three-round debate:

1. Round 1, Declare: each agent states LONG / SHORT / NEUTRAL with thesis text.
2. Round 2, Align: every agent is reclassified as AGREE or DISSENT against the running consensus.
3. Round 3, Verdict: the chair tallies the longs/shorts/neutrals and declares the conviction percentage.

Conviction is converted to a raw probability (50 + score · 40) and then calibrated against the historical hit-rate stored in learning.json. The Settings tab lets you require a minimum calibrated probability (default 46%) and a minimum reward/risk after fees (default 1.2×). A signal must clear both gates plus the conviction floor to be promoted to "actionable".

When a signal becomes actionable the playbook generator (playbook.js) computes a Robinhood-friendly trade plan tuned for tiny accounts:

• Entry: limit order with a small price-protection buffer (no chasing).
• Stop: the larger of 1.0 × ATR, 1.5% of price, or $0.08.
• Target: sized so realized R/R, after round-trip slippage and friction, still meets your minimum (default 1.2×).
• Hold window: never beyond 5 minutes before the regular session close, regardless of intraday targets.
• SHORT picks on Robinhood return a "skip or use a Put option" notice, the app refuses to draw a short order ticket on a cash account.

The sizer uses a risk-based formula with a minimum of 1,000 shares per position:

risk_per_share = stop_distance + 2 × slippage_per_share
          fixed_risk_shares = floor( riskUSD / risk_per_share )
          half_kelly_shares = floor( fixed_risk_shares × 0.5 )
          shares = max( 1, fixed_risk_shares, half_kelly_shares )

Worked example: $4 stock, $0.06 stop distance:

• Slippage per share ≈ $4 × 0.005 × 2 = $0.04
• Risk per share = $0.06 + $0.04 = $0.10
• Fixed-risk shares = floor(25 / 0.10) = 250
• Minimum floor applied → 1,000 shares · order ≈ $4,000 · max loss ≈ $100

Break-even win-rate at the configured R/R floor of 1.2× is 100 / (1 + 1.2) ≈ 45.5%. If your sustained hit-rate falls below that, the system is losing money even when "right".

• UI banner & sound: every actionable signal pops a sheet at the bottom of the screen and a soft chime.
• Browser notifications: granted via the standard permission prompt; titled "marketXED · LONG TICKER".
• SMS: when Twilio credentials are configured in Admin, the same alert is sent to your phone only between 9 a.m. and 4 p.m. ET and de-duplicated for 6 minutes per ticker/direction.

Every alert is recorded in SMS History. There is no auto-execution: you open Robinhood and place the order.

For every actionable pick the system writes a row to predictions.jsonl: ticker, direction, entry, stop, target, conviction, agent votes, and the regime classification. After the trade window closes (or the next day), the actual outcome is matched and the per-agent / per-regime hit-rate is updated.

These rates feed back as multiplicative weights on each agent's confidence in the next cycle. Over time, agents that consistently fail in a given regime are progressively muted; agents that are reliably correct are amplified. There is no neural network and no opaque "AI", every weight is human-readable in learning.json.

• Is the calibrated probability ≥ my Settings threshold?
• Is the after-fees R/R ≥ my Settings threshold?
• Is the spread < 1% of price? (Open the ticker in Robinhood and look.)
• Is there a known earnings, FDA, or macro event inside my hold window? If yes, skip.
• Will the trade put me over the PDT count?
• Is the dollar risk ≤ 1% of total liquid net worth? (Ignore the 25 USD default if your real budget is smaller.)
• Have I had two stops in a row today? If yes, stop trading for the day.
• Did I read the playbook's steps in full and understand each one?

ATR

Average True Range: typical bar-to-bar movement; the basis for stop distance.

R/R

Reward-to-risk: target distance divided by risk distance, after fees.

Conviction

Composite score combining direction strength, agent agreement, and coverage.

Calibrated probability

The raw probability adjusted by the engine's actual historical hit-rate.

Regime

A classification of current market behavior (trending up/down, range, high-vol, etc.).

Veto

A hard rule that disqualifies a signal regardless of score (e.g. liquidity, halts).

Slippage

Difference between the price you saw and the price you got filled at.

Half-Kelly

Half of the Kelly-criterion size: a defensive scaling that survives bad streaks.

EOD

End-of-day: the regular-session 4 p.m. ET close.

PDT

Pattern Day Trader: FINRA designation that restricts day-trading on small accounts.