Astel Research - TradingAgents

Autonomous AI-Powered Cryptocurrency Futures Trading System & Quantitative Signal Research Platform

Astel Research - TradingAgents is an autonomous AI trading system and quantitative research platform combining machine learning predictions, multi-horizon market features, regime-conditioned signal fusion, episodic memory, procedural memory, LLM reasoning, and a self-reflection feedback loop for cryptocurrency futures markets.

The system operates continuously, conducting empirical historical replay, regime-conditioned research, read-only shadow observation, and continuous independent out-of-sample data collection before any execution consideration.

Creator & Lead Researcher: Azis Maulana Suhada
Developed by: PT Authentic Media Services by AMS Capital
Research Areas: Autonomous AI Agents • Quantitative Trading • Machine Learning • Multi-Agent Systems • Large Language Models (LLMs)

---

Current Status

🟢 CURRENT STATUS: COLLECTING INDEPENDENT OOS DATA (PHASE 14B.2)

• CandidateSignalV2: FROZEN (Unmodified since Phase 12.9)
• Phase 14B.1 Protocol: FROZEN (1.0.0-FROZEN, SHA-256: 473a3be30daa2e4820c816169d5111753e3562b0f781cc9269751b3a549d4f36)
• Phase 14B.2 Collector: ACTIVE (agent/_phase14b_oos_collector.py)
• Phase 14C Execution: NOT EXECUTED (Awaiting collection of required 196 raw 4H candles per asset)
• Historical Alpha Claim: NOT ESTABLISHED (Historical replay edge does not constitute independent OOS alpha proof)
• Independent OOS Evidence: INSUFFICIENT until collection completes

Current Operational & Research Stage

The repository is actively executing Phase 14B.2 — Continuous Independent OOS Data Collection.

Independent post-boundary 4H market candlesticks are being continuously accumulated across the canonical 12-asset universe (BTC, ETH, SOL, BNB, XRP, AVAX, LINK, DOGE, ADA, LTC, AAVE, SUI).

All statistical decision thresholds, benchmark definitions, transaction cost models, and CandidateSignalV2 strategy rules remain 100% frozen.

---

Overview

Astel Research - TradingAgents is built on the philosophy that successful autonomous trading requires a multi-layered quantitative and cognitive architecture:

1. Market Perception & Feature Engineering — Engineering multi-horizon returns, oscillators, and volatility metrics to detect market regimes and build probabilistic signals.
2. Quantitative Signal Research — Frozen candidate signal engines evaluated across multi-asset historical datasets and regime switches.
3. Memory Systems — Structured memory across procedural rules, episodic experiences, and counterfactual shadow validation.
4. Reasoning & Intelligence — Multi-agent specialist evaluation, evidence fusion, and LLM-powered contextual reasoning.
5. Execution & Supervision — Isolated, fault-tolerant execution infrastructure with strict risk supervisor controls and shadow branching.

Market Data
    │
    ▼
Feature Engineering
    │
    ├── Machine Learning
    │
    └── Candidate Signal V2 (FROZEN)
              │
              ▼
       Regime Detection
              │
              ▼
       Signal / Conflict
          Resolution
              │
              ▼
      Market Intelligence
              │
              ▼
       Evidence Fusion
              │
              ▼
       Decision Agent
              │
              ▼
       Risk Supervisor
              │
        ┌─────┴─────┐
        │           │
      SHADOW     EXECUTION
        │           │
        ▼           ▼
   Shadow Engine  ExecutionEngine
                     │
                     ▼
                 GateExecutor
                     │
                     ▼
              Gate.io Testnet

Research Safety & Execution Isolation

Candidate V2 evaluation is currently conducted in strictly isolated SHADOW / RESEARCH mode. Candidate V2 does not submit exchange orders (EXECUTION_MODE=SHADOW enforced).

---

Frozen CandidateSignalV2 Strategy Specification

CandidateSignalV2 is FROZEN and has remained completely unmodified since Phase 12.9.

1. Regime Mapping Rules

• BULLISH_TREND → Multi-Horizon Momentum
• BEARISH_TREND → Mean-Reversion
• CONSOLIDATION → Multi-Horizon Momentum
• HIGH_VOLATILITY → NEUTRAL (No trade)

2. Conflict Resolution Rule (FOLLOW_MOMENTUM)

When Mean-Reversion and Momentum indicators generate opposing signals:

• MR_SHORT + MOM_LONG → Follow Momentum LONG
• MR_LONG + MOM_SHORT → Follow Momentum SHORT

3. Indicator Definitions

• Normal Mean-Reversion:
  • LONG: RSI14 < 35 AND Bollinger %B < 0.10
  • SHORT: RSI14 > 65 AND Bollinger %B > 0.90
• Normal Multi-Horizon Momentum:
  • LONG: ret_3 > 0 AND ret_6 > 0 AND ret_12 > 0
  • SHORT: ret_3 < 0 AND ret_6 < 0 AND ret_12 < 0

These parameter thresholds are frozen for out-of-sample validation and must not be tuned or optimized.

---

Historical Research vs. Independent Out-of-Sample Validation

It is a critical quantitative research principle in Astel to strictly distinguish between Historical In-Sample Replay and Independent Out-of-Sample (OOS) Validation.

Historical Replay Dataset (Phases 12.4–14.0)

• Universe: 12 Canonical USDT-margined futures contracts
• Timeframe: 4H
• Sample Size: 1,001 candles per asset (12,012 total bar evaluations)
• Historical Boundary: 2026-10-06T00:00:00Z
• Warmup Window: 90 candles
• Forward Horizons: T+1, T+3, T+6 (Primary Horizon: T+3)

Historical Phase 12.9 / Phase 14.0 Findings

Metric / Scenario	Historical Value
Candidate V2 Gross T+3 Expectancy	+0.2141%
Always LONG Benchmark Expectancy	+0.2073%
Paired Expectancy Difference ($\delta$)	+0.0068% (+0.68 bps)
Paired 95% Bootstrap Confidence Interval	[-0.0693%, +0.0893%]
Probability Difference > 0	58.7%
Optimistic Cost Scenario (0.06% round-trip)	Net +0.1541%, Profit Factor 1.21
Base Cost Scenario (0.14% round-trip)	Net +0.0741%, Profit Factor 1.09
Adverse Cost Scenario (0.25% round-trip)	Net -0.0359%, Profit Factor 0.96

Methodological Conclusion: The historical replay dataset overlapped with the market period used to formulate and freeze Candidate V2. Therefore, these historical results DO NOT constitute independent out-of-sample alpha evidence. CandidateSignalV2 does NOT have statistically proven alpha.

---

Phase 14B.1 — Frozen Independent OOS Protocol

Phase 14B.1 established a machine-readable, immutable validation protocol prior to independent OOS data collection.

Protocol Specifications

• Protocol Version: 1.0.0-FROZEN
• Protocol Manifest SHA-256: 473a3be30daa2e4820c816169d5111753e3562b0f781cc9269751b3a549d4f36
• Historical Boundary: 2026-10-06T00:00:00Z
• OOS Inclusion Rule: new_first_timestamp > 2026-10-06T00:00:00Z
• Universe & Timeframe: 12 canonical assets, 4H timeframe
• Evaluation Window: 90 warmup candles + 100 evaluation observations + 6 horizon safety candles
• Benchmark Set:
  • Always LONG
  • Always SHORT
  • Uniform Random
• Methodological Break Note: The historical Phase 12/14 diagnostic benchmark included REGIME_SWITCH_V1. Phase 14B.1 replaces this with Always LONG, Always SHORT, and Random. Historical comparisons against REGIME_SWITCH_V1 are not directly interchangeable with the frozen Phase 14C benchmark set.
• Statistical Inference: 1,000 paired bootstrap resamples, 95% percentile confidence intervals, deterministic seed 42.
• Trading Cost Scenarios:
  • Optimistic: 0.06% round-trip
  • Base: 0.14% round-trip
  • Adverse: 0.25% round-trip

Protocol Immutability Guarantee: The Phase 14B.1 protocol manifest is frozen and SHA-256 fingerprinted. It must not be modified or adjusted based on post-hoc OOS observations.

---

Phase 14B.2 — Continuous Independent OOS Data Collector

The Phase 14B.2 continuous collector safely and deterministically acquires newly completed 4H candlesticks from Gate.io REST API into an isolated OOS dataset directory (quant_system/data/oos/).

Infrastructure Modules

• Data Collector: agent/_phase14b_oos_collector.py
• Health & Status Validator: agent/_phase14b_oos_health.py

Operational Properties

• Supported Modes: --once (single cycle) or --watch --interval 3600 (hourly continuous polling).
• Completed-Candle Invariant: now_sec >= candle_start_sec + 14400 (Enforces zero lookahead).
• Strict Invariants: UTC timestamps, 4H step alignment, strict monotonicity, rejection of duplicates and overlapping historical batches, atomic .tmp file persistence.
• Fail-Closed Design: Rejects malformed API responses; never executes Phase 14C automatically.

---

OOS Dataset Readiness Requirements

To satisfy the frozen Phase 14B.1 protocol, each canonical asset requires:

$$\text{Raw Completed 4H Candles Required} = 90 \text{ Warmup} + 100 \text{ Evaluation} + 6 \text{ Horizon Safety} = \mathbf{196 \text{ Candles}}$$

Important Data Distinction: $196$ is the RAW COMPLETED CANDLE REQUIREMENT per asset. It MUST NOT be conflated as $196$ independent statistical evaluation observations. The actual evaluation sample size per asset is $N=100$.

Initial collection baseline: 12 total candles (1 candle per asset, 0 usable evaluation observations). The system is NOT eligible to execute Phase 14C until the complete 196 candle threshold is satisfied across all 12 canonical assets.

---

Phase 14C-Prep.1 — Statistical Robustness & Methodology Audit

Phase 14C-Prep.1 was a strictly read-only quantitative methodology audit of the statistical protocol prior to Phase 14C.

Key Audit Findings & Corrected Power Analysis

Using first-principles analytical $Z$-test calculations ($\sigma = 0.02$, two-sided $\alpha = 0.05$):

True Paired Mean Edge ($\delta$)	$N=100$ (Nominal Eval N)	Hypothetical Statistical $N=196$ Scenario	Illustrative $N_{\text{eff}} \approx 50$ Scenario
$0.02%$ ($2$ bps)	$5.11%$	$5.22%$	$5.06%$
$0.05%$ ($5$ bps)	$5.71%$	$6.42%$	$5.36%$
$0.10%$ ($10$ bps)	$7.90%$	$10.97%$	$6.45%$
$0.20%$ ($20$ bps)	$17.01%$	$28.71%$	$11.13%$

Power Analysis Takeaways:

1. Under nominal $N=100$, statistical power to detect a $20$ bps ($0.20%$) mean edge is $17.01%$ (not high power).
2. Achieving $80%$ power at nominal $N=100$ requires a mean edge of $\delta = \mathbf{0.560%}$ ($56$ bps per trade).
3. $N=196$ in statistical power tables is explicitly a HYPOTHETICAL STATISTICAL SAMPLE-SIZE SCENARIO, not the raw completed candle count.
4. $N_{\text{eff}} \approx 50$ is an illustrative AR(1)-based sensitivity scenario ($\rho_1 = 0.33$), not an empirical property of CandidateSignalV2 OOS returns.
5. Moving Block Bootstrap CI widening (+18.68% under synthetic AR(1)) represents a controlled simulation stress test demonstrating theoretical dependence sensitivity.

---

Phase 14B.2 Integrity & Phase 14C Dry-Run Audit

A hardening pass verified collector invariants and Phase 14C fail-closed dry-run behavior using synthetic fixtures:

• Collector Invariants: Verified candle completion, timestamp monotonicity, OHLCV validity, canonical asset coverage, gap detection, duplicate rejection, atomic persistence, and restartability.
• Fail-Closed Dry-Run Suite (agent/test_phase14c_dryrun.py): Verified 10 synthetic edge-case scenarios (insufficient $N$, protocol SHA mismatch, historical overlap, duplicate timestamps, missing assets, incomplete candles, valid dry-run, cost sensitivity).
• Determinism Verification: Confirmed exact deterministic equality (RUN_1 == RUN_2) across bootstrap resamples using seed 42.
• Regression Suite: 88/88 tests passing across all repository Phase 14 test suites.

---

Pre-Phase-14C Repository Cleanup Status

A minimal, reversible pre-Phase-14C cleanup was executed:

• 13 redundant artifacts (legacy Phase 8 audit scripts, one-off branding/fix scripts, informal notes, and temporary test DBs) were moved to archive/pre-phase14c/.
• 2 database files (agent/agent.sqlite, agent/test_fusion.sqlite) were retained in place due to default fallback references.
• Documentation report produced: PRE_PHASE14C_CLEANUP_REPORT.md.

---

Next Methodological Milestones

1. Continuous Collection: Continue running agent/_phase14b_oos_collector.py until the canonical 12-asset universe reaches $\ge 196$ completed 4H candles per asset.
2. Health Verification: Periodically inspect dataset integrity using python -m agent._phase14b_oos_health.
3. Readiness Gate Confirmation: Confirm ready_for_phase_14c: true in phase14b_oos_collector_manifest.json.
4. Dataset Freeze: Lock the eligible OOS dataset SHA-256 fingerprint.
5. Phase 14C Execution: Execute Phase 14C out-of-sample validation strictly according to the frozen Phase 14B.1 protocol. Zero strategy tuning or post-hoc protocol modification permitted.

---

Technology Stack

Category	Technology
Language	Python 3.11+
Web Framework	FastAPI + Uvicorn
Data Processing	NumPy, Pandas, SciPy
ML Models	LightGBM / XGBoost
LLM Framework	Ollama (local) • Groq (cloud) • DeepSeek (cloud)
Exchange API	Gate.io Futures API v4
Data Validation	Pydantic v2
Unit Testing	Python Standard unittest (88 Phase 14 Tests PASS)

---

License

MIT License

Copyright (c) 2026 Astel Research - TradingAgents

---

Acknowledgements

• Gate.io — Futures API & Testnet infrastructure
• Ollama — Local LLM inference platform
• LightGBM — Gradient boosting framework
• FastAPI — Web dashboard framework