Ming Chen.
All projects
04QuantLabQuant trading platform2026Live · private sign-in

A Taiwan-stock quant platform, from research to orders.

Data updates, screening, backtesting, optimisation, machine learning and the post-close order list all run on one platform. Claude helps design and review strategies but sees only in-sample data; out-of-sample validation is always done by code.

Visit site
built-in strategies, 22 backed by a paper
28
built-in strategies, 22 backed by a paper
factors with IC and quantile analysis
20
factors with IC and quantile analysis
API routes
75
API routes
weekday run: signals computed and sent
18:30
weekday run: signals computed and sent
Stock research page: candles, monthly revenue, institutional flows and the AI report

01

Problems

  • No subscription

    A self-hosted research environment with no monthly fee, where the data and code stay under my control.

  • No retail broker API

    Cathay Securities offers no retail API, so trading is semi-automated: compute the list, send a notification, place the orders by hand.

  • Backtest bias

    Backtests easily use data that was not yet published (look-ahead bias) or overfit through excessive tuning.

  • Yahoo throttles cloud IPs

    Yahoo rate-limits cloud servers, so adjusted prices are computed in-house from FinMind raw data.

  • Learning while building

    As a newcomer to quant, I needed an AI tutor that can query data, run backtests and explain the results.

02

Architecture

A Next.js front end on the public domain proxies /api to FastAPI on the private network, with DuckDB for storage. Background jobs and the scheduler run inside the API process, so DuckDB is never locked across processes.

Data sourcesFinMind · TWSE · TPEx · FRED · Yahoo fallback
Claude Code CLILoopback MCP tools only
BrowserGoogle sign-in · allow-list
Next.js 16/api proxy (no 30 s rewrite timeout)
FastAPIBacktests · optimisation · ML · AI
DuckDB19 tables · /data volume
SchedulerMon–Fri 18:30
DiscordSignals and alerts

← Swipe to see the whole flow →

Tech stack

Front end
Next.jsReactTailwind CSSTanStack Querylightweight-chartsECharts
Back end & data
PythonFastAPIDuckDBpandasNumPy
Quant & ML
OptunaLightGBM
AI
Claude CodeMCPAnthropic API
Deployment
DockerZeaburDiscord

03

Key features

  • Stock research

    Candles and indicators, monthly revenue, institutional flows, P/E bands and an AI report.

  • Screener

    Rule-based screening over the universe or all listed stocks, one click to a strategy or alert.

  • Strategy studio

    28 built-in strategies and a visual rule builder, or describe a strategy and let Claude draft it.

  • Backtest report

    Equity vs 0050, drawdowns, monthly heatmap, rolling Sharpe, trades on the chart, AI review.

  • Optimisation

    Optuna in-sample search, out-of-sample check, top-10 robustness, Deflated Sharpe.

  • Factors & ML

    IC and quantile returns for 20 factors; walk-forward LightGBM with feature importance.

  • AI researcher & tutor

    Claude proposes and backtests strategies over several rounds; the tutor queries data, runs backtests and reads screenshots.

  • Screenshot to AI

    Alt-click any card, table or chart and ask Claude about it in the side panel.

  • Daily signals & account

    After the close, an order list sized to each strategy's allocation, round and odd lots separated; real fills recorded.

  • Live tracking & alerts

    Alerts when live results lag the backtest or draw down too far; condition alerts fire only on new matches.

04

Technical details

  1. 04.1

    No look-ahead: data counts from its publication date

    Every data point in a backtest is aligned to the date it actually became available, not the period it describes.

    • Monthly revenue uses FinMind's actual publication time, falling back to the 10th of the next month.
    • Quarterly reports are usable only from the day after the legal deadline: Q1 5/15, Q2 8/14, Q3 11/14, annual 3/31.
    • FRED monthly series are shifted by their publication lag; anything published after today is excluded.
    March revenue
    Not yet
    Q1 report
    Not yet
    FRED monthly
    Not yet
    Backtest day
    JanFebMarAprMayJun
    • March revenue：10th of next month (fallback)
    • Q1 report：Day after the 5/15 deadline
    • FRED monthly：Shifted by publication lag

    Grey is the period the data describes; until the backtest day passes the badge, the data does not exist.

  2. 04.2

    Decide at the close, fill at the next open

    The engine simulates how Taiwan stocks are actually traded, with Taiwan's real costs.

    • Fee 0.1425% × broker discount, 0.1% default slippage, 0.3% sell tax (0.1% for ETFs).
    • Only changed weights trade; untradable days defer the order; buys scale down when cash runs short; no margin, no shorting.
    1. 01Close, day tStrategy sets target weights
    2. 02Open, day t+1Trade only changed weights
    3. 03Sell firstFee + tax + slippage
    4. 04Then buyScaled to available cash
    5. 05Mark to closeEquity, drawdown, Sharpe

  3. 04.3

    The AI never sees out-of-sample data

    The AI researcher can keep proposing and backtesting, but only in-sample. Its finalists are run out-of-sample by code, and the Deflated Sharpe ratio discounts the luck of trying many times.

    • Claude backtests through an internal MCP tool, 12 attempts by default, and submits at most 3 finalists.
    • What the AI produces is a JSON-Schema-validated rule set, never code.

    In-sample: Claude can see

    Attempt 1/12

    Backtest through an internal MCP tool (lab_backtest)

    AI can't cross

    Out-of-sample: code only

    • Finalist 1Waiting
    • Finalist 2Waiting
    • Finalist 3Waiting

    Deflated Sharpe discounts the luck of many tries

    Bars are illustrative, not real results.

05

Screens

Backtest report: metrics and equity curve vs 0050
Backtest report: metrics and equity curve vs 0050
AI researcher: conclusions, every attempt and in- vs out-of-sample Sharpe
AI researcher: conclusions, every attempt and in- vs out-of-sample Sharpe
Optimisation: per-trial scatter plots and parameter heatmap
Optimisation: per-trial scatter plots and parameter heatmap

Next project

AI video automation

FrameFlow

Paste a script and get a narrated explainer video with subtitles and camera moves, uploaded to YouTube and backed up to Google Drive.

View project