AI-POWERED FINANCIAL RESEARCH

AiStockCN Research Copilot

Research companies with evidence, not guesses.

Ask a US-company question and trace the answer back to live market observations and deterministic calculations. SEC filing evidence is the next source joining the same workspace.

Live market evidenceDeterministic calculationsEvidence vs inference

Connected financial platform

Live Research Surface

Research systemLive

US equity universe5,336

Evidence modes3

AALAmerican AirlinesAAPLAppleAMDAdvanced Micro DevicesAMZNAmazon

System architecture

A visible chain of evidence

The design favours explicit artifacts and recoverable steps. A reviewer can move from a signal back through its model, feature snapshot and raw inputs without treating the dashboard as the system of record.

IngestBaoStock + AKShare

NormaliseParquet + PyArrow

EngineerTraining + inference

ModelLightGBM profiles

ValidateWalk-forward tests

OperateSignals + paper trades

artifact_contract.ymlread-only example

snapshot: latest_inference
inputs:
  - daily_kline.parquet
  - daily_valuation.parquet
controls:
  schema: validated
  labels: excluded
  reference_freshness: checked
outputs:
  - inference_features.parquet
  - scored_snapshot.parquet

01 · Source integrity

Canonical registry first

The universe and reference coverage are resolved before downstream features are trusted.

02 · Reproducibility

Artifacts over hidden state

Parquet datasets, model files, profiles and metadata make a run inspectable after it finishes.

03 · Operational safety

Observe before control

Read-only views are the default; authenticated controls are isolated behind explicit roles.

Architecture · Delivery · Analysis

One platform, three disciplines

The strongest evidence is not a diagram in isolation. It is the connection between a business requirement, a technical decision and a working operational surface.

Architecture

A deterministic artifact pipeline

Each stage has a clear input, output and runtime boundary. Raw market files, model-ready panels, scored snapshots, model metadata and backtest results remain independently inspectable.

BoundariesLineageData modelling

Delivery

Restartable, observable workflows

Long-running jobs expose progress, state files, failure reasons and logs. Operators can inspect the exact stage that is running before deciding whether to wait, retry or stop.

Batch controlTelemetryRecovery

Business analysis

Decisions translated into controls

Product needs such as freshness, reproducibility and safe execution are expressed as workflow gates, explicit API responses and focused operator views rather than hidden assumptions.

RequirementsAPI contractsAcceptance

Requirements traceability

Needs become acceptance evidence

Every important requirement is connected to an engineering response and something an operator or reviewer can verify.

Business needEngineering responseAcceptance evidence

Research must be reproducibleSeparate training and inference artifacts; persist model metadata and profile configuration.Model view, saved runs and feature importance

Market data may be incomplete or staleTrack registry coverage, reference-data readiness, missing codes and stale reference states.Data summary, batch state and warning surfaces

Backtests must avoid future leakageUse expanding-window walk-forward evaluation and keep forward labels out of inference snapshots.Saved backtest runs and out-of-sample metrics

Execution must remain reviewableSeparate scored intent from broker reconciliation and persist targets, positions, orders and fills.Paper-trading ledger and gateway status

Architecture decision record

Trade-offs made explicit

These decisions keep the system understandable to technical teams, business stakeholders and future maintainers.

ADR-01

Parquet as the workflow contract

Columnar files keep large time-series datasets efficient while making every intermediate artifact portable and inspectable.

ADR-02

Training and inference remain separate

Two feature paths make label boundaries explicit and reduce the risk of accidental future-data leakage.

ADR-03

The control plane reads operational truth

The API reports from artifacts, containers and runtime state instead of creating a competing copy of pipeline state.

ADR-04

Intent is isolated from execution

Scoring produces reviewable targets; a separate daemon reconciles those targets through the paper-trading gateway.

Implemented end to end

Research code is only one part of the product.

PythonFastAPIPandasPyArrowLightGBMscikit-learnTypeScriptNext.jsReactDockerLinuxREST APIs

See the operating system

The case study explains the choices. The live panel shows the evidence.

Review current pipeline telemetry and a public market-data preview, or sign in to explore the complete operator workflow.

Pipeline Telemetry

Idle

Last state update: Aug 10, 2026, 20:59Last code: 688981Remaining: 0Log source: docker

688981 completed, note: reference_cache_stale_until:2026-08-07
Pass 1 finished, completed 5205/5205, remaining for retry: 0
Starting pass 2/5, pending stocks: 0
{
  "finished": true,
  "total_codes": 5205,
  "done_codes": 5205,
  "remaining_codes": 0,
  "state_file": "quant_data/batch_state/all_a_3y_state.json",
  "last_code": "688981"
}
logout success!

Latest Kline Preview · 000001

Visitor Preview

Rows: 821Latest: Aug 10, 2026Date range: Mar 22, 2023 to Aug 10, 2026

date	open	high	low	close	volume
Aug 10	11.18	11.38	11.16	11.29	88,906,014
Aug 7	11.23	11.26	11.1	11.19	88,297,701
Aug 6	11.22	11.28	11.12	11.27	104,634,257
Aug 5	11.41	11.5	11.18	11.25	151,150,993
Aug 4	11.58	11.62	11.42	11.44	122,112,984
Aug 3	11.54	11.66	11.52	11.62	106,085,094
Jul 31	11.5	11.63	11.28	11.63	202,497,895
Jul 30	11.28	11.62	11.18	11.61	277,770,773