Ito 1.1
An autonomous AI trading system on a real exchange, with real time decisions and complete transparency. Here's exactly how it works, what we learned, and what we changed.
🎯 Our Goal
Good investment tools should be accessible to everyone. Today, sophisticated trading strategies, risk management techniques, and market analysis tools are locked behind expensive hedge funds, private wealth managers, and institutional trading desks. Most people face two choices: figure it out alone, or pay high fees for mediocre advice.
We believe AI can change that. Not by replacing human judgment, but by giving everyone access to the same analytical capabilities Wall Street has used for decades. Ito is our first step toward that future.
To build AI systems that help everyday people make better financial decisions. Whether that's managing a retirement portfolio, growing savings, or actively trading. Not by promising magic returns, but by providing intelligent, transparent, and accessible tools.
Imagine opening an app, telling it your financial goals and risk tolerance, and having an AI assistant that monitors markets for opportunities matching your profile, explains what it sees in plain English, suggests trades with clear rationale and risk/reward, helps you manage positions with professional grade techniques, and learns from your preferences over time.
Why we're building in public
Most trading algorithms are black boxes. You give them money, hope for the best, and have no idea what's happening under the hood. We're taking a radically different approach: complete transparency.
Every trade Ito makes, you see it. Every decision, explained. Every win and every loss, public. We're not hiding behind backtests or cherry picked results. Real exchange, real markets, real time.
🔬 Research Methodology
Ito is a research project, not a finished product. Our goal is to develop and validate a system capable of generating consistent returns of 10% or more annually, outperforming passive index investing (the S&P 500 has averaged roughly 10% per year historically). Until we can demonstrate this level of reliability over extended periods, Ito remains experimental. We trade real capital to ensure authentic market conditions, but the primary output is knowledge, not returns.
The core question
Can a large language model, given sufficient market context and clear decision frameworks, make profitable trading decisions? Classical algorithmic trading systems have proven this is possible with rule based approaches. Quantitative funds have generated consistent alpha for decades using statistical models. The question is whether LLMs can do something similar, or better, by reasoning about market conditions in ways that rigid algorithms cannot.
What we observed in Ito 1
Our first live deployment revealed two big problems. First, the model exhibited classic human trading psychology: positions would reach +1% or +1.5% profit, the model would think "it's going higher," then watch them reverse to -1% or worse before closing. In behavioral finance, this is known as the disposition effectThe tendency to sell winning investments too early while holding losing investments too long. First identified by Shefrin & Statman (1985), building on Kahneman & Tversky's prospect theory (1979).[1,2].
Second, and more fundamental: the model was doing everything in a single step. It would receive raw market data, analyze it, and make trading decisions all at once. This created a "contamination" problem where the pressure to make decisions was affecting the quality of the analysis itself.
Key insight: separation of concerns New
The breakthrough came from a simple observation: professional trading desks separate analysis from execution. Analysts produce research. Traders make decisions based on that research. They are different roles with different incentives and, in many jurisdictions, regulatory mandated information barriers[3]. When the same agent does both simultaneously, cognitive biases creep in. The analyst starts seeing patterns that confirm what the trader wants to do.
That observation is the architectural foundation of Ito 1.1: separate the analyst from the trader. Two distinct steps, two distinct prompts, two distinct outputs. The analyst never knows a trade might happen. The trader never sees raw data. Clean separation.
📐 Standard Frameworks
Before explaining our architecture, it's worth reviewing the established knowledge we're building on. The concepts below are well documented in academic literature and industry practice. We are not claiming novelty here; we are applying known principles to a new context.
Expectancy (textbook formula)
Trading profitability is determined by the relationship between win rate, average win size, and average loss size. Standard material from any quantitative trading course:
Where:
W = Win rate (probability of winning trade)
L = Loss rate (1 - W)
Avg_Win = Average profit on winning trades
Avg_Loss = Average loss on losing trades
The key insight from this formula, known to every professional trader, is that you can be profitable with a 40% win rate if your average win is 2.5x your average loss. Optimizing for win rate alone is a common amateur mistake.
Fee impact (basic arithmetic)
Transaction costs compound with trade frequency. Obvious but often underestimated:
At 20 trades/day with 0.09% round trip: 36% monthly
At 5 trades/day: 9% monthly
At 2 trades/day: 3.6% monthly
This explains why high frequency trading requires high frequency edge. Most signals don't have predictive power on short timeframes sufficient to overcome fee drag.
ATR based sizing (Wilder, 1978)
Average True RangeA volatility measure developed by J. Welles Wilder in his 1978 book "New Concepts in Technical Trading Systems." ATR captures price volatility including gaps, making it superior to simple high-low range for position sizing.[4] is the standard volatility measure for position sizing, documented in any technical analysis textbook. It ensures stops are calibrated to each asset's typical price movement:
True Range = max(High - Low, |High - Prev Close|, |Low - Prev Close|)
Stop Distance = Entry Price ± (ATR × Multiplier)
The result is tight stops on low volatility assets and wider stops on high volatility assets. It adapts automatically to changing market conditions.
MACD z-score normalization
Raw MACD values vary dramatically between assets based on price levels. A MACD of 0.5 might be significant for one asset and noise for another. Z-score normalizationA statistical technique that transforms values to show how many standard deviations they are from the mean. Z-score = (value - mean) / standard deviation. Makes signals comparable across different distributions. is a standard statistical technique that makes signals comparable across different distributions:
Calculated over 75 period lookback.
Z-score > 1.5 means signal is 1.5 standard deviations above average.
🏗️ Two-Step Architecture New
The core innovation in Ito 1.1 is splitting the decision process into two completely separate steps. Not just a code refactor. A fundamental change in how the system reasons about markets.
Why two steps matter
In a single step system, the model receives data and must immediately produce a trading decision. That creates several problems. The model might anchor on a trading idea before fully analyzing the data. It might see confirming evidence more easily than disconfirming evidence. It might skip analysis entirely if the conclusion seems obvious.
With two steps, these problems disappear. The analyst has no idea what the trader will do. It just describes what it sees. The trader never touches raw data. It just evaluates the analyst's report. Neither can contaminate the other.
Data flow
Every 5 minutes, the system executes a complete cycle. Market data is collected from Hyperliquid (prices, candles, funding rates, orderbook depth) and news from financial APIs. The raw data goes to the Analyst. The Analyst produces a structured narrative for each symbol. That narrative, along with portfolio state and P&L history, goes to the Trader. The Trader outputs structured JSON decisions. Those decisions are validated against hard constraints and executed.
🔬 Step 1: The Analyst New
The Analyst receives raw market data and produces structured narrative analysis. It does not trade. It does not recommend. It does not predict. It describes observable market structure, sentiment, and price behavior only.
Analysis framework
For each symbol, the Analyst produces five sections. The framework is inspired by institutional research practices but adapted for perpetual futures markets.
1. Global Structure: "The Dog"
What is price actually doing? The Analyst classifies structure as one of three states: Advancing (higher highs and higher lows), Declining (lower highs and lower lows), or Ranging (overlapping, no clear direction). It identifies the current phase: impulse, consolidation, breakout, pullback, or range. Key support and resistance levels come from recent swing points, not from arbitrary percentages.
2. Local Sentiment: "The Tail"
What are perpetual traders positioned for? Funding rateIn perpetual futures, funding is a periodic payment between longs and shorts to keep the perp price close to spot. Positive funding = longs pay shorts (crowded long). Negative = shorts pay longs (crowded short). indicates who is paying whom. Positive funding means longs are paying shorts, suggesting crowded long positioning. Negative funding means the opposite. Order book imbalance and spread provide additional context but are not predictive signals.
3. Dog vs Tail
The key synthesis step. A well known framework"Dog vs Tail" compares what price is actually doing (the Dog / global structure) with what traders are positioned for (the Tail / local sentiment). When they align, trend has conviction. When they diverge, the crowd may be wrong. in institutional trading, "Dog vs Tail" compares what price is doing (global structure) with what traders are positioned for (local sentiment). The Analyst states this explicitly in one line: "Global [Advancing/Declining/Ranging] / Local [Bullish/Bearish/Neutral]" followed by whether this represents alignment or divergence.
4. Reality Check
Does price action confirm, deny, or ignore the prevailing narrative? News is summarized with explicit time decay: fresh (less than 24h), recent (1 to 2 days), or stale (likely priced in). The Analyst assigns one of four states:
- PRICED IN: News is old, price has already moved
- ABSORPTION: News released but price ignoring it
- DIVERGENCE: Price moving opposite to narrative
- IMPULSE RISK: Fresh catalyst, potential for sharp move
5. Recent Price Action
A factual description of the last 6 candles. Candle size, wicks, range expansion or compression. No patterns unless mechanical and obvious. No subjective adjectives like "strong" or "weak".
The Analyst is explicitly forbidden from predicting future price movement, suggesting trades or positioning, using subjective adjectives, or filling gaps with assumptions. If data is missing or unclear, it must say so. That discipline ensures the Trader receives clean, unbiased information.
💼 Step 2: The Trader New
The Trader receives the Analyst's narrative, the current portfolio state, P&L history for open positions, and trading constraints. It never sees raw market data. Its job is to make execution decisions by selecting from a predefined hypothesis menu.
The Hypothesis Menu
Instead of generating trades from scratch, the Trader must match each situation to one of five predefined hypotheses. That constraint improves consistency and makes decisions auditable. The hypothesis framework is standard in systematic trading; what's novel is using it as a menu for an LLM.
| Hypothesis | When to Use | Sizing |
|---|---|---|
| A: Trend Continuation | Clear HH+HL or LH+LL, pullback to structure | 10 to 15x |
| B: Mean Reversion | Ranging structure, price at range extremes | 5 to 8x |
| C: Sentiment Fade | Dog vs Tail divergence (fade the crowd) | 5 to 10x |
| D: Breakout Play | Consolidation + IMPULSE RISK catalyst | 5 to 8x |
| E: No Trade | No clear edge, mixed signals, PRICED IN | None |
Hypothesis E is the default. The Trader must actively justify departing from "do nothing". In Ito 1, the system would often force trades when flat. Now, E is the path of least resistance. Most cycles end with E for most symbols.
Invalidation conditions
Every entry requires an explicit invalidation condition from the Analyst's structure. Example: "MACD crosses bearish OR price closes below $450 support". If the invalidation triggers, the position closes. If it doesn't, the position holds. No time based exits. No fixed profit targets. That approach, standard among professional discretionary traders, was the major change from Ito 1's rule based exits.
P&L Memory New
The Trader sees the complete P&L history of each open position: current P&L, peak P&L reached, time since peak, and recent trajectory. That addresses the disposition effect by making the Trader explicitly aware of profits being given back. If a position reached +1.5% and is now at +0.3%, that information is front and center.
Current P&L: +0.3%
Peak P&L: +1.5% (reached 45 min ago)
P&L Trend: falling
Drawdown from Peak: -1.2%
Original Invalidation: "MACD crosses bearish"
// The Trader must check: has invalidation triggered?
// If YES then close_position
// If NO then hold (even though giving back profit)
Output format
The Trader outputs strict JSON. For each symbol: signal (buy_to_enter, sell_to_enter, hold, close_position), hypothesis (A through E), confidence (0 to 1), leverage, stop_loss, take_profit, invalidation_condition, and a public message explaining the decision in plain language.
🛡️ Behavioral Safeguards New
Beyond the two-step architecture, Ito 1.1 implements multiple hard coded behavioral rules that cannot be overridden by the AI. They address specific failure modes observed in testing.
Anti-churn rules
Overtrading is the fastest path to losses. These rules are enforced at the engine level:
After closing any position, no new entry on the same symbol for 30 minutes. Prevents emotional re-entry.
No entry within 0.5% of last exit price. Prevents churn around the same level.
Maximum 1 trade per symbol per hour. Forces patience and conviction.
Maximum 1 new position per 5-minute analysis cycle. Quality over quantity.
Correlation exposure limits
To prevent concentration risk from correlated positions:
- Symbols are grouped by correlation (tech stocks together, index together)
- Maximum 2 positions in the same direction within a correlation group
- Prevents scenarios like: MSFT short + AMZN short + TSLA short + NDX short
Gain protection rules New
These rules trigger automatically based on P&L trajectory, addressing the disposition effect mechanically:
Level 1: PnL >= +1.0% then Stop moves to breakeven +0.1%
Level 2: PnL >= +1.2% then Take 20% partial exit (quick lock)
Level 3: PnL >= +2.0% then Take 25% partial exit
Level 4: PnL >= +3.5% then Take 33% of remaining
Trailing: PnL >= +2.5% then 0.6% trailing stop activates
// Emergency tightening:
Peak PnL >= +0.8% AND drawdown >= 0.3% then Stop moves to breakeven
These thresholds are calibrated for perpetual futures with typical 5 to 15x leverage. The key insight: we don't trust the AI to take profits. We force it mechanically.
Hard daily limits
| Constraint | Limit | Rationale |
|---|---|---|
| Max entries/day | 8 | Fee drag control |
| Max entries/hour | 2 | Pacing control |
| Max concurrent positions | 5 | Concentration for quality |
| Max daily loss | 10% | Circuit breaker |
🔍 Independent Risk Monitor New
The trading engine runs every 5 minutes to analyze markets and make decisions. But what happens if the engine is slow, crashes, or misses a rapid price move? To solve this, we run a separate risk monitoring process every 60 seconds that operates independently from the main trading logic.
In production trading systems, risk management should never depend on the same code path as trade execution. If the analysis engine hangs on an API call or crashes mid-cycle, positions are still protected. The risk monitor uses database locks to prevent conflicts, ensuring exactly one process touches positions at any time.
What the risk monitor does
Every 60 seconds, the risk monitor wakes up, acquires a database lock, and performs three operations:
Fetches current prices from Hyperliquid and updates all open positions with real time P&L calculations.
Checks if any position has hit its stop loss or take profit. If triggered, closes the position immediately.
Applies hard coded rules to tighten stops and take partial profits. These rules execute automatically and cannot be overridden.
Hard protection rules
These rules trigger based on P&L trajectory and are enforced at the risk monitor level, not by the AI. They execute every minute, independent of what the Trader thinks:
IF PnL > +1.0% AND distance to TP < 0.6%
THEN stop moves to breakeven +0.1%
// Rule B: Peak drawdown protection
IF peak PnL >= +0.8% AND drawdown from peak >= 0.3%
THEN stop moves to breakeven +0.1%
// These rules are checked every 60 seconds
// They only tighten stops, never loosen them
The key principle: stops can only get tighter, never looser. If the risk monitor calculates a new stop that would be worse than the current one, it keeps the current stop. This prevents any scenario where protection is accidentally removed.
Partial exit cascade
The risk monitor also manages the scaled exit system. Each position tracks its "exit level" (0 to 4), and the monitor advances through levels as P&L thresholds are reached:
| Level | Trigger | Action |
|---|---|---|
| 0 → 1 | PnL reaches +1.0% | Stop moves to breakeven +0.1% |
| 1 → 2 | PnL reaches +1.2% | Take 20% partial exit (quick lock) |
| 2 → 3 | PnL reaches +2.0% | Take 25% partial exit |
| 3 → 4 | PnL reaches +3.5% | Take 33% of remaining |
| Any | PnL reaches +2.5% | 0.6% trailing stop activates |
Exit levels are one way. Once a position reaches level 2, it cannot go back to level 1 even if P&L drops. The partial exits are already locked in. This prevents the system from repeatedly taking the same partial exit if price oscillates around a threshold.
Database locking
Both the main trading engine (every 5 min) and the risk monitor (every 1 min) can modify positions. To prevent race conditions, they share a MySQL advisory lock:
SELECT GET_LOCK('ito_engine_lock_v1', 0) AS got_lock
// If got_lock = 0, another process is active
// Skip this cycle and try again next minute
// On shutdown:
SELECT RELEASE_LOCK('ito_engine_lock_v1')
The lock timeout is 0 seconds (non-blocking). If the lock is held, the process exits immediately rather than waiting. This ensures we never have two processes fighting over the same positions.
Both processes check for an emergency stop flag file on startup. If the file exists, they exit immediately without touching any positions. This gives us a manual kill switch that works even if the database is unreachable.
⚖️ Risk Management
Risk management follows standard quantitative trading practices. These parameters are not novel; they are industry consensus values. What's different is how they integrate with the two step architecture.
Position level controls
| Parameter | Value | Standard Range |
|---|---|---|
| Max risk per trade | 2% of capital | 1% to 3% typical |
| Max position size | 15% of capital | 10% to 25% typical |
| Stop loss distance | ATR based | Volatility adjusted |
| Take profit distance | ATR based | 2:1 R minimum |
Position sizing formula
We use standard ATR based position sizing[4]:
Where:
Max Risk USD = Account Balance × 2%
Stop Distance = |Entry - Stop Loss|
Margin Required = (Position Size × Entry Price) / Leverage
Capped at: 15% of Account Balance
Liquidation protection
For leveraged perpetual futures, we calculate liquidation price assuming 5% maintenance margin:
Short: Liquidation = Entry × (1 + 1/Leverage - 0.05)
Stop losses are always placed well before liquidation price. Typically at 1 to 3% distance versus 5 to 20% to liquidation depending on leverage.
🔄 Iteration Cycle
Our research follows a systematic iteration process: deploy, observe, analyze, hypothesize, test. Each version represents a specific hypothesis being tested against live market conditions.
Single step architecture. Identified disposition effect and analysis contamination problems.
Two step architecture (Analyst + Trader). Hypothesis menu. P&L Memory. Invalidation based exits. Behavioral safeguards.
Refinements based on observed behavior. Focus areas to be determined.
Only after demonstrating reliable 10%+ annual returns across multiple market conditions.
What we measure
Success criteria
The research will be considered successful when Ito demonstrates consistent profitability (positive returns over rolling 3 month periods), risk adjusted performance (Sharpe ratio above 1.0), drawdown control (maximum drawdown below 20%), and market outperformance (returns exceeding S&P 500 over comparable periods). Until these criteria are met across multiple market cycles, Ito remains a research project.
❓ Frequently Asked Questions
Is this on a real exchange?
Yes. Ito trades with $10,000 on Hyperliquid, a real decentralized exchange. We use real capital because paper trading doesn't capture execution realities like slippage, fees, and psychological pressure. But this is research capital, not an investment product.
What changed in Ito 1.1?
The core change is the two step architecture: separate Analyst and Trader modules that cannot contaminate each other. The Analyst describes market structure without knowing trades might happen. The Trader makes decisions without seeing raw data. We also added the hypothesis menu (five predefined trade types), P&L Memory (tracking profit history), invalidation based exits (no more time stops), and hard coded behavioral safeguards.
What is Dog vs Tail?
A framework used in institutional trading. "Dog" is what price is actually doing (the global structure). "Tail" is what traders are positioned for (local sentiment via funding rates). When they align, the trend has conviction. When they diverge, the crowd may be wrong. The Analyst explicitly reports this alignment or divergence.
What is the hypothesis menu?
Instead of generating trades freely, the Trader must select from five predefined hypotheses: A (trend continuation), B (mean reversion), C (sentiment fade), D (breakout), or E (no trade). That constraint improves consistency and makes decisions auditable. E is the default; most cycles end with no trade on most symbols.
How does gain protection work?
We implement mechanical rules that trigger based on P&L trajectory: stop moves to breakeven at +1%, partial exits at +1.2%/+2%/+3.5%, and trailing stop at +2.5%. These rules cannot be overridden by the AI. They execute automatically. That addresses the disposition effect (holding winners too long until they reverse).
When will Ito be ready for real use?
When we can demonstrate consistent 10%+ annual returns with acceptable drawdowns across multiple market conditions. That could take months or years. Until then, this is research, not a product.
Is this financial advice?
No. Ito is an experimental research project. Nothing shown here constitutes financial advice. Past performance does not guarantee future results. Trading involves risk of loss. Do your own research.
References
- Kahneman, D., & Tversky, A. (1979). Prospect Theory: An Analysis of Decision under Risk. Econometrica, 47(2), 263-291.
- Shefrin, H., & Statman, M. (1985). The Disposition to Sell Winners Too Early and Ride Losers Too Long: Theory and Evidence. The Journal of Finance, 40(3), 777-790.
- FINRA Rule 2242 (Debt Research Analysts and Debt Research Reports). Research conflict management and separation requirements.
- Wilder, J. W. (1978). New Concepts in Technical Trading Systems. Trend Research.
Follow the Research
Watch Ito trade live, see every decision explained, and follow along as we test hypotheses and iterate.