hardy30894/market-screener：面向 AI 与半导体板块的选股与退出监控工具

A discovery-first, earnings-aware stock screener for the AI / memory / semiconductor / photonics complex, with an exit monitor, a market-regime dashboard, and a real backtest that tells you whether any of it actually works.

It scans the whole US tech-hardware market live, ranks coiled pre-breakout setups, flags strong names reporting soon, and watches your holdings for exit signals. It runs itself daily in the cloud and writes one report.

It is a decision-support tool. It scans the market, ranks pre-breakout setups, flags strong names reporting soon, and watches your holdings for exits. For an honest read of what it can and cannot do, see What this is NOT and the backtest.

Table of contents

• What it does
• What this is NOT
• Components
• Quick start
• Usage
• How the score works
• Does it actually work? (backtest)
• Data reliability
• Automation
• Limitations and roadmap
• Disclaimer

What it does

1. Discovers across the entire US tech-hardware universe (semis, equipment, photonics, components, plus optional solar and quantum), not a hand-picked list. Names you have never heard of surface on their own.
2. Scores each survivor 0 to 100 by stacking fundamental-inflection signals (estimate-revision momentum, revenue acceleration, margin expansion) with technical pre-breakout signals (proximity to highs, trend, momentum, volatility contraction).
3. Flags earnings: every row shows days-to-earnings, and a dedicated EARNINGS WATCH block surfaces strong picks reporting soon, front and center.
4. Monitors exits: feed it your holdings and it returns HOLD / TRIM / EXIT with the reasons firing (trend break, estimate cuts, stop levels).
5. Reads the regime: a risk-on / risk-off dashboard from yield curve, credit spreads, breadth, volatility, and semis leadership.
6. Validates itself: a point-in-time backtest measures whether the factors predict forward returns, out of sample, after costs.

Every morning it assembles all of this into a single dated report.

What this is NOT

This section exists because the easiest thing to build is a screener that looks authoritative and is quietly wrong. This one tries hard not to be that.

• Not a validated alpha engine. The backtest says the technical factors carry weak signal (IC_IR about 0.47 at a quarterly horizon, t-stat 1.44, below significance) and a top-quintile basket roughly reproduced the semis ETF. Treat scores as a research queue, not buy signals.
• Not a crash predictor. The regime dashboard reads ambient risk. It has high false-positive rates and long, variable lead times. Nobody predicts crashes, and this does not pretend to.
• Not a backtested earnings or exit strategy. Earnings tags and exit rules are sensible and literature-backed (post-earnings drift, trailing stops), but they are signals and risk-management discipline, not proven-on-this-data edges.
• Not reliable data, yet. It runs on yfinance (an unofficial Yahoo scraper). The data layer is hardened (cache, retry, fail-loud) but yfinance is still the ceiling. See Data reliability.

If a number looks confident, check its cov (coverage) column and remember the backtest verdict.

Components

Quick start

pip install yfinance pandas numpy

python screener.py                 # discover + score the tech-hardware market
python screener.py --all-themes    # add solar + quantum
python monitor.py NVDA MU COHR     # exit signals on holdings
python regime.py                   # market-risk dashboard
python backtest.py                 # does the strategy actually work?
bash daily_report.sh               # the full combined report

Usage

Screener

python screener.py                       # AI/memory/semis/photonics (default)
python screener.py --growth 15           # only names already growing rev >15%
python screener.py --all-themes          # everything incl. solar + quantum
python screener.py --quantum --seeds NOK # add quantum pure-plays + specific tickers
python screener.py --min-score 60        # only strong setups
python screener.py NVDA MU CRDO          # score exactly these

Output splits into CORE (large-cap) and EARLY (small/mid), each ranked, with a plain-English verdict (primed, early (fundies), extended, watch, broken trend) and an EARNINGS WATCH block at the top.

Exit monitor

python monitor.py --file holdings.txt    # one ticker per line

Regime + backtest

python regime.py
python backtest.py --start 2016-01-01 --horizon 63

How the score works

Three layers, stacked. The edge (such as it is) comes from combining them, not from any single filter.

1. Fundamental inflection (leading). Forward-EPS revision momentum (the highest-weighted factor, best-documented predictor), revenue-growth acceleration, gross-margin expansion, last earnings surprise.
2. Technical pre-breakout (timing). Proximity to the 52-week high (coiled, not extended), 12-1 momentum, trend (50 over 200), relative strength vs the SOX, volatility contraction.
3. Liquidity gate. Price and dollar-volume floors to keep junk out.

The technical weights are calibrated to measured backtest rank-IC; the fundamental block is theory-backed but, on free data, an untested overlay (see below). Full detail in SCREENER.md.

Does it actually work? (backtest)

backtest.py measures whether the factors predict forward returns, point-in-time, out of sample, after costs. On about 78 names over 2016 to 2026, quarterly horizon:

Honest verdict: weak signal at a quarterly horizon, not statistically significant (t-stat 1.44, below the 2.0 bar), and the long top-quintile basket returned about 36% CAGR vs the SOX ETF's 35.5%. In plain terms: the technical machinery roughly reproduces the sector beta. The fundamental factors (the ones that probably carry the real edge) are excluded from the backtest because free data is not point-in-time. Use this as a discovery and timing tool, not a return generator.

Data reliability

The screener talks only to a DataSource interface, so the provider is a one-file swap. The yfinance implementation is hardened:

• Cached to disk (6h), so reruns do not re-hammer Yahoo.
• Retried with exponential backoff on transient errors.
• Fails loud: if more than 25% of fetches fail or more than 40% of data is stale, the run aborts with exit code 2 and prints why, instead of scoring on garbage. Silent-wrong is the failure mode that hurts; this turns it into a visible stop.

The ceiling: yfinance is still an unofficial scraper with no SLA. The real fix is a paid API key (FMP / Tiingo / Polygon), which also unlocks point-in-time data and a genuine backtest. The abstraction is ready for that swap.

Automation

A GitHub Actions workflow runs the full report every day and commits it to results/, so it is hands-off and independent of any local machine. Read the latest at results/latest.md.

Limitations and roadmap

• Validated edge is weak. The single highest-value upgrade is a point-in-time paid data feed, which would both fix reliability and let the fundamental factors (and exits, and regime) be properly backtested. That is the path from "weak technical edge" to a real, measured edge.
• Survivorship bias in the backtest (current listings only) flatters the absolute returns; read the spread, not the level.
• yfinance reliability is capped until the API swap.
• backtest.py still calls yfinance directly and is not yet behind the data layer.

Disclaimer

This is software for research and education. It is not investment advice, not a recommendation to buy or sell any security, and carries no warranty. All forward-looking figures referenced in the macro thesis are forecasts subject to revision. Markets carry risk, including loss of principal. Do your own research and consult a licensed professional before acting on anything here.