How AI Day Trading Bots Actually Work: The 8-Stage Pipeline from Data to Execution

How Does an AI Day Trading Bot Work?

An AI day trading bot is a software system that autonomously screens markets, evaluates trade candidates using machine learning and large language models, sizes positions based on risk parameters, executes orders through a broker API, and manages exits in real time. The best systems also learn from outcomes, retraining models on actual trade results to improve future decisions. Here is the 8-stage pipeline behind Tradewink's Paper Autopilot. Every stage below runs in a simulator or a paper/sandbox account. Public subscriptions are paper-only; separately approved private beta accounts may submit live broker orders.

Stage 1: Pre-Scan Gates

Before scanning a single ticker, the system runs safety checks to determine whether trading conditions are favorable. This prevents the bot from trading during dangerous or low-probability periods.

Market Regime Detection

A Hidden Markov Model (HMM) classifies the current market regime by analyzing SPY returns. The model identifies three states:

• Bull regime: Trending up, momentum strategies favored
• Sideways regime: Range-bound, mean reversion strategies favored
• Bear regime: Trending down, defensive positioning, reduced size

The regime classification determines which strategies are active, how aggressively the system sizes positions, and whether certain trade types are skipped entirely.

Intraday Regime Overlay

On top of the daily HMM regime, a 5-minute efficiency ratio on SPY classifies intraday conditions as "trending," "choppy," or neutral. If the intraday regime flips from trending to choppy mid-session, the system triggers an AI-powered exit debate on open positions.

Monk Mode

Monk mode is a pre-scan gate that skips trading during unfavorable conditions:

• First and last 15 minutes of regular hours (plus a midday dead-zone block, 11:30–14:00 ET, when enabled)
• Regime transitions (when the HMM regime has recently changed)
• Pre-earnings blackout (default: 2 days before earnings)
• Low-conviction periods (when recent trade outcomes have degraded confidence)

Stage 2: Ticker Screening

The screener narrows a liquid universe (about 180 static names plus Finviz/heatmap additions) to at most 25 candidates per scan using quantitative filters.

Ticker Sources

Candidates come from three pools, merged and deduplicated:

1. Default universe: about 180 high-liquidity tickers (AAPL, NVDA, TSLA, META, etc.), plus a micro-account universe of ~50 cheaper names
2. Finviz dynamic sourcing: Real-time top gainers, unusual volume, gap-ups/downs
3. User watchlist: Tickers added by the trader get priority scanning and a +15 point score boost

Scoring System

Each ticker is scored on a 0–150 composite (not 0–100) across weighted factors such as:

Only tickers scoring above the minimum threshold proceed to strategy evaluation.

Stage 3: Strategy Evaluation

The strategy engine runs each screened ticker through multiple strategy algorithms simultaneously. Tradewink uses 16+ strategies across three categories:

Momentum Strategies

• Breakout: Price above resistance on elevated volume
• Gap and Go: Pre-market gap with continuation volume
• Opening Range Breakout (ORB): First 15-minute range expansion

Mean Reversion Strategies

• RSI Bounce: Entry when RSI drops below 30 with volume confirmation
• VWAP Reversion: Price deviation from VWAP with mean-reversion entry
• Bollinger Band Squeeze: Entry at lower band with contracting bandwidth

Advanced Strategies

• Options Flow Confirmation: Unusual options activity confirming directional bias
• Pairs Trading: Cointegrated pairs with z-score divergence
• Sector Rotation: Capital flow between sectors based on relative strength

Each strategy produces a signal with direction (long/short), entry zone, stop-loss, target price, and a raw score.

Signal Discretization

Raw strategy signals are classified into a 5-tier system: Strong Buy, Buy, Neutral, Sell, Strong Sell. An ML-based signal quality classifier further evaluates whether the signal pattern historically led to profitable trades.

Stage 4: AI Conviction Scoring

This is where modern AI trading systems differ from classical algorithmic trading. Instead of relying solely on quantitative signals, the system uses large language models to evaluate each candidate with contextual reasoning.

Default: Single LLM Call

For each candidate that passes screening and strategy evaluation, a single AI call generates a conviction score from 0 to 100. The model receives:

• Technical setup data (indicators, levels, volume)
• Fundamental context (earnings date, insider activity, sector performance)
• Market regime classification
• Recent trade outcomes for similar setups

Conviction is an additive, capped boost on the composite score — not a 1.15×/0.80× multiplier. By default, conviction under 60 rejects the candidate; 60–79 adds a small boost and 80+ adds up to +8 points.

Opt-In: Multi-Agent Team Evaluation

For deeper analysis, a 3-agent team evaluates up to a few top candidates per scan:

1. Technical Analyst
2. Risk Analyst
3. Execution Strategist Disagreement is a hard veto (score to 0). Bull/bear debate is used separately for some exit decisions, not as this entry team.

This multi-agent debate catches risks that quantitative filters miss — like an upcoming Fed meeting that changes the macro backdrop, or a sector rotation pattern that makes the trade thesis weaker.

Model Routing

AI calls are routed through OpenRouter with explicit per-tier model IDs (not a published Claude/GPT-5/Gemini roster). A model router selects a lighter or heavier model per task. This keeps AI costs manageable even with hundreds of daily evaluations.

Stage 5: Position Sizing

Position sizing determines how much capital to allocate. The system calculates three sizes and uses the most conservative:

Risk-Based Sizing

Standard percentage-of-equity risk. With a $10,000 account and 1% risk per trade, maximum dollar risk is $100. If the stop-loss is $2 away from entry, position size = 50 shares.

ATR-Based Sizing

Uses Average True Range to normalize for volatility. A volatile stock (high ATR) gets fewer shares than a stable stock, holding dollar risk constant. Multiplier: 1.5x ATR for day trades, 2x ATR for swing trades.

Half-Kelly Sizing

The Kelly Criterion calculates optimal bet size based on win rate and reward/risk ratio. Half-Kelly reduces the theoretical optimum by 50% to account for estimation error and reduce variance.

Micro Account Mode

For accounts under $1,000, the system automatically enables:

• Fractional shares (typically two decimal places on brokers that support them)
• Reduced risk per trade (3% max vs 1-2% standard)
• Lower concentration limit (25% of equity per position)
• $1 minimum order value

The most conservative of the three methods wins. This means position sizes are always at or below the optimal risk level.

Stage 6: Trade Execution

Once a candidate is sized, it passes through the execution pipeline:

Risk Manager Gate

The risk manager checks:

• Daily loss limit: If the day's losses exceed the configured threshold, no new trades
• Position count: Maximum concurrent positions not exceeded
• Sector concentration: Not overexposed to any single sector
• Day-trade / margin checks: Round trips may still be counted for display, but the federal PDT designation ended June 4, 2026 and Tradewink's PDT gate no longer blocks entries. Brokers may apply transitional intraday-margin rules through October 20, 2027
• Circuit breaker: If a cascade of losses occurs in a short window, trading pauses

Broker Submission

Paper orders are filled in Tradewink's simulator or submitted to a connected paper/sandbox broker account through an abstract broker interface. Public plans do not include live order submission. The system is non-custodial.

Each paper trade is logged with an audit trail: timestamp, strategy, conviction score, sizing rationale, paper confirmation, and simulated fill price.

Smart Execution

A VWAP/TWAP slicer exists in code but is off by default (smart_execution_enabled = false). Typical day-trade orders are submitted without that slicing.

Stage 7: Dynamic Exit Management

Open positions are monitored continuously with multiple exit mechanisms running in parallel:

Trailing Stops with Paper-Account Sync

As a paper position moves in your favor, the stop-loss ratchets upward. The system cancels the old paper stop order and submits a new one at the trailed level — so the paper stop always matches the system's calculated level.

MFE/MAE Tracking

Maximum Favorable Excursion (MFE) and Maximum Adverse Excursion (MAE) are tracked on every tick. This data feeds into the ML exit model and post-trade analytics.

ML-Driven Exit Engine

A machine learning model evaluates 6 possible actions on each monitoring cycle:

• Hold position
• Tighten stop
• Take partial profits
• Exit fully at market
• Extend target
• Hold or exit (the stock day-trade pipeline does not reverse into shorts)

The model is trained on historical trade outcomes, learning which exit patterns maximize realized P&L for each strategy type.

Regime-Shift Exit

If the intraday regime flips (e.g., trending to choppy), the system triggers a bull/bear AI debate on whether to hold or exit. This catches regime transitions that purely quantitative stops would miss.

Time-Based Exit

Positions that haven't moved after 60 minutes by default (configurable; mean-reversion max hold 120 minutes) are closed at market. Stagnant positions tie up capital with opportunity cost. A separate zero-MFE rule can exit even earlier if the trade never showed traction.

End-of-Day Flatten

All day trade positions are closed before market close. No overnight risk on day trades.

Stage 8: Self-Improvement Loop

The most underappreciated component of an AI trading system is the feedback loop. Without it, the system's performance degrades as market conditions evolve.

Post-Trade Reflection

After every closed trade, a TradeReflector AI analyzes:

• What went right or wrong
• Whether the entry, sizing, and exit were optimal
• Lessons learned for similar future setups

These reflections are stored in a database and injected into future conviction scoring calls, giving the AI memory of past mistakes.

Walk-Forward ML Retraining

Every 1.5 hours, ML models retrain on recent trade outcomes using walk-forward validation. This prevents overfitting to stale data while adapting to current market conditions.

UCB-Tuned Strategy Selector

A reinforcement learning bandit (UCB-Tuned by default; Thompson Sampling is an option) maintains a belief over each strategy's performance by regime. Strategies that perform well in the current regime get higher weight; underperformers get reduced allocation. This happens automatically — no manual strategy selection needed.

Confidence Calibration

The system tracks whether its confidence scores match actual outcomes. If the AI consistently rates trades at 80% confidence but they only win 55% of the time, the calibrator adjusts future scores downward until predicted and actual performance align.

Architecture Diagram

Pre-Scan Gates
  |-- HMM Regime Detection (SPY)
  |-- Intraday Regime Overlay (5-min efficiency ratio)
  |-- Monk Mode (quiet hours, earnings blackout, regime transitions)
  v
Screening (~180 static tickers + Finviz + watchlist)
  |-- Volume, ATR%, gap, RSI, relative volume, liquidity, catalyst scoring (0–150 composite)
  v
Strategy Evaluation (16+ strategies)
  |-- Momentum, mean reversion, breakout, flow, pairs
  |-- Signal discretization (Strong Buy -> Strong Sell)
  v
AI Conviction Scoring
  |-- Single LLM call (default) or 3-agent debate (opt-in)
  |-- Model routing: cheapest appropriate model per task
  v
Position Sizing
  |-- Risk-based, ATR-based, half-Kelly (most conservative wins)
  |-- Micro account adjustments for <$1,000 accounts
  v
Execution
  |-- Risk manager gate (daily loss, concentration, circuit breaker; PDT gate is a no-op)
  |-- Paper submission via abstract BrokerClient (simulator or paper/sandbox account)
  v
Dynamic Exits
  |-- Trailing stops synced to the paper account
  |-- ML exit model (6 actions)
  |-- Regime-shift debate, time-based exit, EOD flatten
  v
Self-Improvement
  |-- Trade reflection (AI-generated lessons)
  |-- Walk-forward ML retraining (every 1.5 hours)
  |-- UCB-Tuned strategy selector (Thompson Sampling optional)
  |-- Confidence calibration

What AI Day Trading Bots Cannot Do

No AI trading system can predict the future or guarantee profits. Here's what to be realistic about:

• Black swan events: No model predicts sudden crashes, halts, or geopolitical shocks. Risk management (stop-losses, position limits) is the only defense.
• Regime changes: When the market shifts from bull to bear, historical patterns break down. The system adapts via retraining, but there's always a lag.
• Slippage and fills: In fast-moving markets, your fill price may differ from the AI's target. Smart execution helps but can't eliminate slippage entirely.
• Overfitting risk: ML models can learn patterns that don't generalize. Walk-forward validation mitigates this but doesn't eliminate it.
• Costs: Commissions, slippage, and AI API costs reduce gross returns. Position sizing must account for total transaction costs.

The CFTC has explicitly warned that "AI won't turn trading bots into money machines." Any platform promising guaranteed returns is a red flag.

Frequently Asked Questions

How much money do I need to start AI day trading?

Tradewink's public offering is paper trading only, so no money is needed to use it. Its paper engine supports micro-account settings (equity under $1,000) with fractional shares and adjusted risk parameters. FINRA's $25,000 PDT minimum ended June 4, 2026; some brokers may still apply PDT-style limits during a transition through October 20, 2027. Cash accounts still need settled funds, and a U.S. margin account still needs $2,000 minimum equity. Tradewink itself does not block a "fourth day trade."

Does the AI trade with my money or its own?

Neither. Tradewink's public offering is paper trading only — it trades simulated money in a simulator or a paper/sandbox account. Public plans do not include live order submission, and it never holds, transfers, or has custody of your funds. For public-plan users, any real trade is your own decision, placed by you at your broker.

How many trades does the system make per day?

It depends on market conditions. In a trending bull market, the system may find 5-10 opportunities. In choppy or quiet markets, monk mode may skip trading entirely. Quality over quantity — the system only trades when the pipeline produces high-conviction setups.

Can I override the AI's decisions?

Yes. Automation is fully optional. You can use Tradewink in signal-only mode (receive AI trade ideas and decide yourself) or configure Paper Autopilot with your own risk limits, excluded tickers, and preferred strategies. Either way, Tradewink stays paper-only.

Is this a black box?

No. Every signal comes with a full written analysis explaining the reasoning, technical setup, entry/stop/target levels, conviction score, and strategy type. The software is MIT-licensed and source-available on request; the public site does not ship the full trading engine as a self-hosted product.