AITradra

Evidence-first AI market intelligence, systematic research and fail-closed trading automation

Market evidence → Mythic multi-agent research → Systematic validation → Plugin ablation → Risk veto → Empirical precision → Protected execution

Product · Architecture · Systematic Research · Plugin Accuracy · Benchmark Gate · Safety · Quick Start

Important

AITradra does not guarantee profit, investment returns, benchmark outperformance or any future accuracy percentage. All accuracy, precision, backtest and benchmark figures are evidence measurements under specific historical or paper conditions. Paper trading is the safe default and funded execution remains fail-closed.

Product

AITradra is an open-source AI-native market intelligence and trading engineering platform designed to answer a harder question than “what does the AI think?”:

Does the strategy, model or plugin add measurable out-of-sample value after costs, risk controls and benchmark comparison?

The platform combines source-aware market research, one authoritative multi-agent decision pipeline, systematic strategy discovery, independent event-driven validation, real FinBERT sentiment, portfolio optimization, forward shadow evidence and strict execution qualification.

Core capabilities

Capability	Purpose
Market intelligence	Price action, OHLCV, news, macro and source provenance
Mythic multi-agent research	Technical, fundamental, macro, sentiment, sector, catalyst and risk reasoning
Real FinBERT	`ProsusAI/finbert` financial-news sentiment classification
Systematic strategy engine	Momentum, SMA trend, mean reversion and breakout candidate discovery
Statistical robustness	Train/validation/test, walk-forward, bootstrap, null tests, multiple-testing adjustment and regime stability
Backtrader replay	Independent event-driven replay with fees, slippage and long/short/exit transitions
Beat-the-benchmark gate	Requires OOS performance relative to SPY, NIFTY 50 or BTC depending on market
Plugin Ablation Lab	Measures whether optional plugins improve hit rate or probability calibration
Shadow paper ledger	Tamper-evident forward decision evidence without execution authority
Portfolio optimization	Half-Kelly, volatility controls and PyPortfolioOpt HRP for multi-asset allocation
Empirical precision gate	Uses resolved directional outcomes and Wilson statistical bounds
Risk Manager	Deterministic pre-order veto and sizing constraints
Strict live-price contract	Primary real-time source plus independent verifier; no provider fallback
Protected broker routing	Paper, Hyperliquid and explicit CCXT venue routing
Health evidence	Safety CI, frontend build, smoke checks and exact-SHA health ledger

What changed in the trading-accuracy upgrade

The current main branch removes several sources of false confidence that commonly appear in AI trading systems.

FinBERT is now actually FinBERT. The sentiment agent runs ProsusAI/finbert; it does not silently replace the model with a general LLM while continuing to call the output FinBERT.
Vibe strategy backtests must contain measured metrics. Empty or zero-only output cannot be classified as a successful validation result.
Vibe Swarm no longer receives a fixed predictive confidence. Successful plugin execution is not treated as proof of market accuracy.
Optional plugins must earn positive influence. FinBERT, Quantic and Swarm outputs are measured through forward ablation before they can receive positive weighting.
Quantic disagreement can reduce risk before it earns positive weight. This keeps external validation asymmetric and conservative.
PyPortfolioOpt is genuinely active. HRP is used when multi-asset histories are supplied, while central per-position safety caps remain enforced.
CCXT routing is explicit and fail-closed. A configured venue is selected deliberately; the system does not silently redirect a funded order to another broker.
The old competing legacy decision chain is retired. Legacy orchestration is now compatibility-only and routes analysis into the authoritative Mythic pipeline.
Core vectorized research uses NumPy/Pandas. Backtrader remains the independent event-driven validator; AITradra does not claim vectorbt is powering core research when it is not.

Architecture

AITradra now has one authoritative research-to-decision direction:

flowchart TD
    U[User / Ticker]

    U --> Q[Query Router]
    Q --> DATA[Research Data + RAG + News + OHLCV]
    DATA --> MYTHIC[Mythic Orchestrator]

    MYTHIC --> TECH[Technical / Fundamental / Macro / Sector / Catalyst]
    MYTHIC --> FIN[Real FinBERT]
    MYTHIC --> OPT[Optional Quantic / Swarm Validators]

    TECH --> SIG[Signal Aggregator]
    FIN --> SIG
    OPT --> SIG

    SIG --> SHADOW[Shadow Forward Ledger]
    SHADOW --> ABL[Plugin Ablation Lab]
    ABL -. measured policy .-> SIG

    DATA --> SYS[Systematic Research Engine]
    SYS --> ROB[Train / Validation / Untouched Test]
    ROB --> STAT[Walk-forward / Bootstrap / Null Test / Regime]
    STAT --> BT[Backtrader Replay]
    BT --> BM[Benchmark Gate]

    SIG --> RM[Risk Manager]
    BM --> RM

    LIVEPRICE[Strict Decision-grade Live Price] --> RM
    RM -->|BLOCK| STOP[No New Entry]
    RM --> QF[Trade Qualification Firewall]

    PREC[Resolved Precision Evidence] --> AUDIT[Tamper-evident Precision Audit]
    AUDIT --> QF
    BM --> QF
    AUTH[Explicit Execution Authorization] --> QF

    QF -->|PAPER| PAPER[Paper Broker]
    QF -->|HYPERLIQUID| HL[Hyperliquid]
    QF -->|EXPLICIT CCXT| CCXT[Binance / Bybit / OKX]
    QF -->|BLOCK| STOP

Execution authority boundaries

Layer	Can research?	Can influence confidence?	Can authorize funded execution?
Research specialists	Yes	Yes	No
FinBERT	Yes	Only after measured ablation policy	No
Quantic / Swarm	Yes	Only under measured policy; disagreement may reduce	No
Systematic Research Engine	Yes	Produces strategy evidence	No
Backtrader validation	Yes	Produces deployment evidence	No
Benchmark scorecard	Yes	Can block strategy eligibility	No
Shadow ledger	Yes	Supplies forward evidence	No
Risk Manager	No	N/A	Can veto only
Precision gate	No	N/A	Can block only
Qualification firewall	No	N/A	Final permission boundary
Broker adapter	No	No	Executes only after permission

See docs/RESEARCH_EXECUTION_BOUNDARY.md for the detailed contract.

Systematic strategy research

core/systematic_research.py performs lightweight deterministic candidate discovery before event-driven replay.

Strategy families

Current catalog includes:

Momentum — multiple lookback and threshold combinations
SMA crossover — fast/slow trend combinations
Mean reversion — rolling z-score entry configurations
Breakout — multiple breakout windows

Validation sequence

Candidate discovery is deliberately separated from final validation:

60% training window — screen candidate strategies.
20% validation window — rank top candidates without touching the final test window.
20% untouched test window — challenge the selected winner.
Walk-forward stability — require acceptable performance across sequential periods.
Block bootstrap — estimate Sharpe robustness under resampled return blocks.
Sign-flip null test — challenge whether observed positive performance could be random directional noise.
Multiple-testing adjustment — penalize strategy selection across many candidate trials.
Regime stability — examine performance across trend and volatility regimes.
Composite robustness score — combine independent robustness evidence.
Backtrader replay — independently replay dated BUY / SELL / EXIT transitions with trading friction.
Benchmark comparison — compare OOS performance against the relevant market benchmark.

Default systematic settings are configurable through .env:

SYSTEMATIC_RESEARCH_ENABLED=true
SYSTEMATIC_RESEARCH_MIN_HISTORY=260
SYSTEMATIC_RESEARCH_MAX_CANDIDATES=40
SYSTEMATIC_RESEARCH_TOP_K=5
SYSTEMATIC_RESEARCH_MIN_ROBUSTNESS_SCORE=60
SYSTEMATIC_RESEARCH_MIN_TEST_SHARPE=0.25
SYSTEMATIC_RESEARCH_MAX_TEST_DRAWDOWN_PCT=25
SYSTEMATIC_RESEARCH_MIN_TRIAL_ADJUSTED_PROBABILITY=0.70
SYSTEMATIC_RESEARCH_MAX_SIGNFLIP_P_VALUE=0.20
SYSTEMATIC_BOOTSTRAP_SAMPLES=400
SYSTEMATIC_SIGNFLIP_SAMPLES=400

Note

A strategy passing systematic research is eligible for further qualification, not guaranteed profitable and not automatically authorized for a funded order.

Beat-the-benchmark gate

A strategy is no longer considered deployable merely because its isolated backtest is profitable.

The event-driven out-of-sample period is compared against a broad default benchmark:

Traded market	Default benchmark
US / general equities	`SPY`
Indian NSE symbols ending in `.NS`	`^NSEI` — NIFTY 50
Crypto	`BTC-USD`

For DEPLOY eligibility, the strategy must satisfy the existing full-sample and OOS validation gates and pass the benchmark scorecard criteria, including:

higher OOS total return;
higher OOS Sharpe ratio;
drawdown not materially worse than the benchmark, with a small absolute tolerance for near-flat benchmarks.

The benchmark scorecard can also calculate:

active return;
annualized alpha;
beta;
information ratio;
directional hit rate;
Brier score when probability forecasts are available;
regime-level strategy vs benchmark performance.

Benchmark results are historical/paper evidence only and are not a promise of future outperformance.

Plugin accuracy & ablation

self_improvement/plugin_ablation.py answers a critical question:

Did this plugin improve the decision compared with the same core system without it?

The lab evaluates optional plugin output against resolved shadow decisions using the same underlying outcomes.

Plugin policies

Policy	Meaning
`KEEP`	Sufficient forward evidence shows incremental value
`ADVISORY`	Insufficient or inconclusive evidence; no positive confidence credit
`DISABLE`	Measured evidence shows material calibration regression without directional benefit

Primary metrics:

directional hit-rate delta;
core Brier score;
plugin-blended Brier score;
probability calibration improvement;
resolved sample count.

By default a plugin needs a meaningful number of resolved samples before it can move from ADVISORY to positive influence.

Current behavior

FinBERT: actual financial sentiment model; positive influence only after measured ablation evidence.
Quantic / SMC: unproven agreement does not automatically boost confidence; meaningful disagreement may reduce confidence as a conservative veto.
Vibe Swarm: advisory synthesis; successful execution does not manufacture predictive confidence.

This design makes plugin count irrelevant. Only measured incremental contribution matters.

Real FinBERT sentiment

agents/sentiment_classifier.py loads:

ProsusAI/finbert

The model processes financial headlines and aggregates positive, negative and neutral probabilities into a bounded sentiment score and confidence value.

Important behavior:

lazy model initialization;
CUDA is used when available;
CPU fallback is supported;
model failure returns neutral/unverified evidence rather than silently switching model identity;
the output records the verified model name;
positive trading influence is still governed by forward plugin ablation.

FinBERT therefore provides a real domain-specific sentiment source without being treated as automatic alpha.

Shadow forward evidence

self_improvement/shadow_trade_store.py records actionable decisions before they have permission to become funded trades.

Each shadow decision stores:

ticker;
direction;
confidence;
entry price;
strategy identifier;
evaluation horizon;
core probability snapshot;
plugin snapshot;
evidence metadata;
timestamp;
SHA-256 chained row hash.

Later, self_improvement/shadow_resolver.py resolves due decisions using market prices and appends outcome evidence.

The original prediction row is not rewritten. This allows AITradra to measure:

forward directional hit rate;
average directional return;
plugin incremental contribution;
calibration quality;
tamper-evident evidence-chain validity.

Shadow evidence has:

execution_authority = false

It exists to improve measurement, not to bypass trading controls.

Portfolio optimization

The Portfolio Agent keeps central risk limits authoritative while supporting two levels of sizing.

Single asset

Uses:

Half-Kelly sizing;
historical win/loss ratio;
annualized volatility sizing;
configured central MAX_POSITION_PCT cap.

Multi asset

When asset_price_history is supplied, AITradra attempts PyPortfolioOpt Hierarchical Risk Parity (HRP).

The resulting weights are then constrained by the same central per-asset cap. If HRP cannot run, the system uses a bounded inverse-volatility fallback rather than inventing weights.

Portfolio optimization improves allocation and diversification engineering; it does not create predictive edge by itself.

Authoritative multi-agent pipeline

AITradra previously contained both a legacy 14-agent LangGraph chain and the newer Mythic orchestration path. That could allow different endpoints to produce different decision processes.

The current architecture makes MythicOrchestrator authoritative.

The legacy AgentOrchestrator remains only as a compatibility facade for older callers and routes analysis into the same Mythic research path.

Current direction:

QueryRouter
   ↓
Data / RAG / News / Intelligence Snapshot
   ↓
MythicOrchestrator
   ↓
Specialists + Critique + Optional Validators
   ↓
SignalAggregator
   ↓
Risk / Research / Benchmark / Precision Qualification

QUICK, DEEP and INSTITUTIONAL research modes therefore share one authoritative decision architecture instead of competing legacy conclusions.

Strict real-time data contract

Qualification-sensitive pricing is handled by gateway/live_price_session.py.

No execution-provider fallback

The first enabled real-time market-data connection is the authoritative provider.
Its observation is valid only inside the configured freshness window.
After expiry, the same authoritative provider must provide a fresh observation.
If the authoritative provider fails, AITradra blocks instead of silently switching providers.
Historical SQLite data, stale cache and research-source fallbacks cannot become decision-grade execution prices.

Default validity:

LIVE_PRICE_VALIDITY_SECONDS=120
LIVE_PRICE_REQUIRE_CROSSCHECK=true
LIVE_PRICE_MAX_CROSSCHECK_DIFF_PCT=1.0

A second independently configured provider is used only to validate the primary quote. It is never promoted into a fallback execution source.

The decision path blocks when the primary price is missing, stale, unusable, unverifiable or materially inconsistent with its independent cross-check.

Trading safety

Safe defaults

PAPER_TRADE_MODE=true
AUTOTRADE_ENABLED=false
MANUAL_LIVE_TRADING_ENABLED=false
REQUIRE_PROTECTIVE_ORDERS=true
REQUIRE_STRATEGY_VALIDATION=true
REQUIRE_EMPIRICAL_PRECISION_VALIDATION=true

Main controls

paper trading by default;
separate manual-live and autonomous-live permissions;
explicit live acknowledgement;
central position-size limits;
daily-loss breaker;
cash reserve requirement;
maximum open-position control;
leverage cap;
stop-loss / take-profit geometry validation;
protective-order enforcement;
reduce-only close support;
existing-position add-ons disabled by default;
systematic + Backtrader strategy validation;
out-of-sample benchmark gate;
current-signal confidence threshold;
empirical precision threshold;
tamper-evident precision audit;
strict live-price freshness gate;
independent live-price disagreement gate;
explicit broker selection;
secret scanning in Safety CI.

Adding credentials does not enable autonomous funded trading.

Broker routing

Paper broker

The default execution route. It requires a market reference price and models:

adverse slippage;
trading fees;
cash balance;
long equity-style positions;
persistent trade records in the broader practice system.

Hyperliquid

The primary specialized crypto execution integration. Live signing remains behind explicit execution authorization.

CCXT

AITradra includes explicit CCXT routing for configured exchanges such as:

Binance;
Bybit;
OKX.

A CCXT venue must be explicitly configured. The router does not silently substitute another funded broker if that venue is unavailable.

Empirical precision gate

Displayed model confidence is not treated as measured accuracy.

Default autonomous-live evidence settings include:

AUTOTRADE_TARGET_PRECISION=0.99
AUTOTRADE_MIN_SIGNAL_CONFIDENCE=90.0
AUTOTRADE_MIN_EVALUATED_SIGNALS=100
AUTOTRADE_MIN_PRECISION_LOWER_BOUND=0.95
PRECISION_LOOKBACK_DAYS=90
PRECISION_VALIDATION_MAX_AGE_DAYS=30

A new autonomous live entry requires, among other controls:

an actionable non-HOLD signal;
Risk Manager approval;
valid protective levels;
explicit live authorization;
current signal confidence above the configured minimum;
approved and fresh strategy validation;
successful benchmark-relative validation where applicable;
enough resolved directional observations;
observed precision at or above the configured target;
Wilson lower-bound evidence at or above the configured threshold;
recent evidence;
a valid tamper-evident precision chain.

Warning

A configured 99% target is an intentionally strict eligibility condition. It is not a statement that AITradra is currently or will be 99% accurate.

Practice trading

Practice mode models:

real/reference market prices when available;
adverse slippage;
fees;
cash and positions;
realized and unrealized P&L;
protective exits;
persistent practice state.

Defaults:

PAPER_STARTING_BALANCE=100000
PAPER_SLIPPAGE_BPS=5
PAPER_FEE_BPS=4

Paper performance depends on the assumptions used and is not guaranteed to match funded trading.

Operations & health evidence

AITradra uses measured workflow status rather than static “all systems green” documentation claims.

Safety CI

.github/workflows/safety-ci.yml runs on pushes to main, pull requests and manual dispatch.

It currently verifies:

critical Python compilation;
embedded-secret checks;
data-validation regressions;
primary-source retry behavior;
trading-safety controls;
empirical precision and precision-audit logic;
strict live-price session behavior;
self-improvement evidence handling;
customer/live-integrity regressions;
systematic research regressions;
portfolio risk-cap behavior;
FinBERT/Vibe/plugin-accuracy regressions;
shadow ledger and benchmark scorecard regressions;
broker-routing regressions;
production React build.

The workflow records the exact tested SHA in the GitHub job summary. Artifact uploads are deliberately disabled to avoid unnecessary Actions storage use.

Live System Smoke

The live-system smoke workflow exercises public-network and application paths while forcing safe execution settings:

PAPER_TRADE_MODE=true
AUTOTRADE_ENABLED=false
MANUAL_LIVE_TRADING_ENABLED=false

It does not submit funded trades.

Health ledger

Workflow outcomes are also published into:

Issue #41 — automated CI and live-smoke health ledger

Use the workflow badges, GitHub Actions run and exact-SHA ledger entry as the measured source of truth.

Quick start

Requirements

Python 3.12+
Node.js 22+
Git

1. Clone

git clone https://github.com/logeshv586-code/AITradra.git
cd AITradra
cp .env.example .env

Keep the safe defaults unless you are intentionally validating a controlled live environment:

PAPER_TRADE_MODE=true
AUTOTRADE_ENABLED=false
MANUAL_LIVE_TRADING_ENABLED=false

2. Backend

python -m venv venv

Linux/macOS:

source venv/bin/activate

Windows PowerShell:

.\venv\Scripts\Activate.ps1

Install dependencies and start:

pip install -r requirements.txt
python main.py

Note

The FinBERT model is loaded lazily the first time sentiment classification is required. Initial model loading can require additional memory and download time.

3. Frontend

cd ui
npm ci
npm run dev

Production build:

npm run build

4. Decision-grade live market data

Research/public-data functionality can run separately. Qualification-sensitive live trading requires configured decision-grade real-time market-data connections.

For the default cross-check policy, configure two independent providers: the first is authoritative and the second verifies it.

Testing

Focused safety and trading-accuracy suite:

python -m pytest -q \
  tests/test_data_validation.py \
  tests/test_primary_source_retry.py \
  tests/test_trading_safety.py \
  tests/test_precision_gate.py \
  tests/test_precision_audit.py \
  tests/test_strict_live_price_session.py \
  tests/test_self_improvement.py \
  tests/test_customer_experience.py \
  tests/test_live_integrity.py \
  tests/test_systematic_research.py \
  tests/test_portfolio_risk_cap.py \
  tests/test_trading_plugin_accuracy.py \
  tests/test_shadow_benchmark_portfolio.py \
  tests/test_broker_routing_accuracy.py

Frontend:

cd ui
npm ci
npm run build

Project structure

AITradra/
├── agents/
│   ├── orchestrator.py                 # Authoritative Mythic research orchestration
│   ├── query_router.py                 # Research-mode and ticker routing
│   ├── sentiment_classifier.py         # Real ProsusAI/finbert sentiment
│   ├── signal_aggregator.py            # Evidence-aware signal fusion
│   ├── strategy_generator_agent.py     # Vibe strategy generation + metric validation
│   └── legacy/                         # Compatibility agents; legacy orchestration redirects to Mythic
├── brokers/
│   ├── broker_router.py                # Explicit paper / Hyperliquid / CCXT routing
│   └── hyperliquid_broker.py
├── core/
│   ├── systematic_research.py          # Candidate discovery and statistical robustness
│   ├── portfolio_optimizer.py          # HRP + bounded fallback allocation
│   ├── trading_safety.py               # Strategy evidence and qualification support
│   └── config.py                       # Central safety and research configuration
├── gateway/
│   ├── server.py
│   ├── live_price_session.py           # Strict decision-grade price contract
│   └── knowledge_store.py
├── self_improvement/
│   ├── plugin_ablation.py              # KEEP / ADVISORY / DISABLE plugin measurement
│   ├── shadow_trade_store.py           # Tamper-evident forward decision ledger
│   ├── shadow_resolver.py              # Later-price outcome resolution
│   ├── benchmark_scorecard.py          # Strategy vs benchmark evaluation
│   ├── precision_store.py
│   └── engine.py
├── tests/                              # Safety, accuracy, benchmark and routing regressions
├── ui/                                 # React + Vite application
├── scripts/                            # Smoke and evidence utilities
├── docs/                               # Architecture and design documentation
└── .github/workflows/                  # Safety CI and live-system validation

Safety position

Area	Repository policy
Research-source provenance	Preserved and separated from execution pricing
Systematic candidate selection	Train/validation/untouched-test separation
Backtest realism	Independent Backtrader replay with fees and slippage
Benchmark comparison	Required for current strategy deployment eligibility
Optional plugin confidence	Must be earned through measured forward ablation
Shadow evidence	Tamper-evident and non-executing
Decision-grade live price	Strict primary provider path
Execution-provider fallback	Disabled
Independent price verification	Validation only; never substitutes primary
Stale/cache execution substitution	Blocked
Portfolio sizing	Central cap remains authoritative
Risk Manager veto	Enforced before qualification
Protective orders	Required by default
Empirical precision	Required by default for autonomous live entries
Practice execution	Default
Autonomous funded execution	Disabled by default
Manual funded execution	Disabled by default

What AITradra does not claim

AITradra does not claim that:

more agents automatically increase accuracy;
more plugins automatically increase accuracy;
FinBERT guarantees better returns;
a backtest proves future profitability;
an AI confidence score equals a measured win rate;
passing a benchmark historically guarantees future alpha;
a 99% configured precision threshold means the system currently achieves 99% accuracy;
paper fills reproduce all live market conditions.

The platform is designed to measure, reject and audit these assumptions instead of hiding them.

Roadmap

Current next-stage research directions include:

larger cross-sectional and multi-asset strategy families;
nested walk-forward re-optimization;
richer factor and statistical-arbitrage research;
liquidity and turnover-aware portfolio optimization;
broader benchmark families;
long-horizon shadow evidence dashboards;
automatic plugin policy visualization;
deeper calibration analysis by market regime;
Strategy Lab UI for systematic robustness and benchmark evidence.

New strategies and plugins should enter as research candidates first, then earn influence through measured evidence.

Community

Contributions are welcome when they preserve the core principles of:

No fake data.
No silent execution fallback.
No confidence presented as empirical accuracy.
No research component granting itself trading permission.
No plugin receives positive weight without evidence.
No profitability guarantee.
Every funded path remains fail-closed.

Before opening a PR:

python -m pytest -q
cd ui && npm run build

Prefer focused tests for every new trading, plugin, broker, benchmark or safety behavior.

License

MIT License. See LICENSE.

AITradra

Research broadly. Validate independently. Measure forward. Execute only when qualified.

Name		Name	Last commit message	Last commit date
.github		.github
agents		agents
alerts		alerts
autoresearch		autoresearch
brokers		brokers
core		core
data		data
docs		docs
gateway		gateway
graphify-out		graphify-out
ingestion		ingestion
llm		llm
mcp		mcp
memory		memory
scheduler		scheduler
scrapers		scrapers
scratch		scratch
scripts		scripts
self_improvement		self_improvement
skills		skills
tests		tests
tmp		tmp
tools		tools
ui		ui
.env.example		.env.example
.gitignore		.gitignore
AITRADRA_MASTER_PLAN.md		AITRADRA_MASTER_PLAN.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
ROADMAP.md		ROADMAP.md
SECURITY.md		SECURITY.md
__init__.py		__init__.py
docker-compose.yml		docker-compose.yml
k3s.bat		k3s.bat
main.py		main.py
market_rag.py		market_rag.py
mcp_config.json		mcp_config.json
move_explainer.py		move_explainer.py
package-lock.json		package-lock.json
requirements.txt		requirements.txt
scratch_test.html		scratch_test.html
simulation_status.json		simulation_status.json
verify_chat_discovery.py		verify_chat_discovery.py

Latest commit

History

Folders and files

Repository files navigation

AITradra

Evidence-first AI market intelligence, systematic research and fail-closed trading automation

Product

Core capabilities

What changed in the trading-accuracy upgrade

Architecture

Execution authority boundaries

Systematic strategy research

Strategy families

Validation sequence

Beat-the-benchmark gate

Plugin accuracy & ablation

Plugin policies

Current behavior

Real FinBERT sentiment

Shadow forward evidence

Portfolio optimization

Single asset

Multi asset

Authoritative multi-agent pipeline

Strict real-time data contract

No execution-provider fallback

Trading safety

Safe defaults

Main controls

Broker routing

Paper broker

Hyperliquid

CCXT

Empirical precision gate

Practice trading

Operations & health evidence

Safety CI

Live System Smoke

Health ledger

Quick start

Requirements

1. Clone

2. Backend

3. Frontend

4. Decision-grade live market data

Testing

Project structure

Safety position

What AITradra does not claim

Roadmap

Community

License

AITradra

About

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages