An autonomous equity screener that gates on a backtested macro model before it spends a dollar of attention, hunts multi-year outperformers printing fresh all-time highs, verifies its own work, and feeds every run back into its next version. It runs unattended three times a market day — and it gets sharper the longer it runs.
Most screeners are a dumb filter: point them at the market, get a list. This one starts by asking a different question — should I even be buying today? A logistic-regression macro model reads the yield-curve slope, the change in Fed Funds, the VIX, and short-term momentum, and emits a single bullish/bearish probability. On risk-off tape it sits out by design and posts an alert. Only when the regime is favorable does the screen run.
When it runs, it pulls every stock printing a new 52-week high across the NYSE, NASDAQ, and CBOE, then squeezes them through two quantitative passes: +100% or more over the last year, then +500% or more over ten years, still within a whisker of its all-time high, still trending. Survivors are the >5x/10y list — companies that have compounded for a decade and are still making new highs.
Then the agent does the part most tools skip: it checks its own work. It re-verifies a random sample of the stocks it rejected, flags any candidate whose gains came from a single suspicious quarter or a sub-$2 penny-stock origin, and compares today's hits against the last run to catch silent regressions. Results go to Telegram and to this page; new names are synced into the live buy-and-hold portfolio sheet. No human in the loop.
LogReg model on yield curve, Fed Funds, VIX & momentum decides bullish vs. bearish. Bearish → alert and stop.
Pull every new 52-week high on NYSE / NASDAQ / CBOE (IPO > 1yr) from Finviz.
Download 2yr quarterly OHLC, keep only names up ≥ 100% over the last year.
Download 10yr history; require ≥ 500%, a fresh all-time high, and positive momentum.
Forbidden-proxy guards, random spot-check re-verification of rejects, regression check vs. the prior run.
Post results to Telegram and this page; append new names to the buy-and-hold portfolio sheet.
Most “AI tools” are frozen the day they ship. This one isn’t — it is built around a loop that treats its own track record as training data.
Andrej Karpathy’s framing of modern AI — Software 2.0, and the Tesla “Data Engine” flywheel — is that you don’t hand-write the rules. You specify a goal, then let an optimization loop improve the system by closing the gap between what it predicted and what actually happened. It is the same loop that trains a neural network — predict, measure the error, adjust, repeat — except run continuously, in production, where every real-world outcome becomes the next lesson. This screener is that loop wearing a finance hat:
The macro model scores the regime; the funnel emits candidate hits. 3× every market day.
Every run appends a labeled record to run_log.jsonl — signal, macro inputs, funnel counts, hits.
Backtest accuracy & Sharpe; live regression alerts and spot-checks are the production error signal.
Model coefficients and every threshold live in config, re-fit from the accumulated log — not hand-coded.
The tuned model ships and runs again. Each cycle adds another labeled example to the flywheel.
The payoff is concrete. The 50-day moving-average filter the screener started with was a static rule someone guessed. It was replaced by a logistic-regression model fit on seven years of data — lifting backtested Sharpe from 0.57 to 0.71 and turning a hand-picked threshold into eight learned coefficients that any future run can re-fit. That is the difference between a script and an agent: a script does the same thing forever; an agent uses its own production logs to write its next version.
Auto-published by the agent on every run — the same output that goes to the Telegram channel. The macro signal, the funnel, and the candidate stats are public; the live symbols are a subscriber signal.
No names cleared the funnel this run.
A logistic-regression classifier trained in our GPD research framework on 2018–2025 market data with walk-forward cross-validation. It replaced a single hand-set moving-average rule — the first turn of the optimization loop above.
| Model | Logistic Regression (sklearn) |
| Backtest window | 2018–2025 · 7yr walk-forward CV |
| Directional accuracy | 56.7% (p < 0.0001) |
| Sharpe ratio | 0.71 vs 0.57 B&H |
| CAGR | 11.8% vs 9.8% B&H |
| Max drawdown | −33.5% vs −35.7% B&H |
| Decision rule | score > 0 → run screen |
| Fallback | SPY > 50-day MA if data unavailable |
This page documents an autonomous research and screening system operated by CSI Automation. Screen output is a quantitative signal, not investment advice and not a solicitation. Backtested results are hypothetical, reflect a specific model over a specific window, and do not predict future returns. Live candidate symbols are withheld; nothing here is a recommendation to buy or sell any security.