← Back to terminal
Methodology

How the ArbibX AI picks stocks

Most AI stock tools tell you what to buy. This page tells you howthey got there — the data, the models, the screens, and the things the system can't see. Read this before you act on a pick.

The pipeline

From 11,000 tickers to 15 picks, in four stages

01 · Universe

Pre-screen the market

Each refresh starts with the top ~250 most-liquid US tickers across mega/large/mid-cap names — a curated universe drawn from Polygon's daily aggregates. Penny stocks, OTC names, and anything with broken data are excluded up front.

02 · Screen

Trim to ~30 candidates

Quantitative filters narrow the universe to roughly thirty names the model is allowed to rank. Filters look at recent price momentum, RSI extremes, distance from SMA20/SMA50, volume vs. its 30-day average, ATR (volatility), 52-week levels, and proximity to upcoming earnings. The goal is to feed the AI a high-signal short list rather than the whole market.

03 · Analyse

Claude does the heavy lifting

Each candidate's full quantitative profile + recent news headlines are sent to Anthropic's Claude. We use a model fallback chain — Opus 4.7 (default) → Sonnet 4.6 → Haiku 4.5 — so a temporary Opus rate-limit just downgrades the analysis instead of breaking it. Claude is prompted as a senior quantitative analyst and must cite specific numerical evidence for every recommendation.

04 · Rank

Output the Top 15

Claude returns a structured JSON array: 15 picks with a signal (Strong Buy / Buy / Hold / Sell), a confidence score, a target price, a one-line thesis, and the key risk. We sanity-check the JSON, snapshot the result to the Track Record table, and serve it to the front end.

Signal cheat-sheet

What Strong Buy / Buy / Hold / Sell actually mean

Strong Buy

High-conviction setup. Multiple positive signals stacked (momentum + volume + favourable news + technical break). Target price typically 8-15% above current.

Buy

Constructive setup with at least one strong tailwind. Reasonable risk/reward over the next 1-4 weeks. Target price typically 4-8% above current.

Hold

Neutral. Either the signals contradict or the stock has already priced in the obvious catalysts. Not a sell — just not a fresh buy here.

Sell

Negative bias. Used when momentum is rolling over, RSI is overbought into resistance, or news flow has turned. Communicates risk to existing holders.

Important:these are short-horizon read-the-tape calls, not long-term recommendations. Confidence is the model's self-rated conviction, not a probability of being right.

Sources

Where the data actually comes from

Polygon.ioDaily/intraday OHLCV bars, snapshot quotes, 52-week levels, volume — the price truth source.
Polygon NewsRecent headlines per ticker so Claude can weigh sentiment, not just price action.
Polygon ReferenceSector/industry classification, name + ticker mapping, fundamentals like market cap.
Anthropic ClaudeThe actual ranking + thesis writing. Currently Opus 4.7 (1M ctx) with Sonnet/Haiku fallback.
Cadence + limits

What this system can & can't do

Strength

Spot multi-signal setups fast

The screener narrows ~250 names to a tight short list, and Claude reads the technicals + headlines together — that's hours of manual work in seconds.

Strength

Show its work

Every pick comes with the numerical evidence Claude cited. You can drill into the stock detail view and verify the indicators yourself.

Strength

Track itself honestly

The Track Record tab keeps every historical Buy/Strong Buy pick — winners and losers — and benchmarks the average against the S&P 500.

Limit

Predict short-term price moves

It can't. Markets are noisy. Treat picks as research starting points, not forecasts.

Limit

Read your inbox or social media

The model only sees the data we feed it. Headlines yes, your Discord chatter no.

Limit

Account for taxes, fees, or your situation

Position sizing, capital gains, timing for your tax year — none of that is in the prompt. That part is on you.

FAQ

Questions we get a lot

How often does the Top 15 refresh?

Up to once every couple of hours during US market hours, and once overnight. The 'Run fresh analysis' button on the Top 15 tab lets you force a refresh; the result is cached for everyone so we don't burn through API quota with one-off requests.

Is this financial advice?

No. ArbibX is a research and screening tool — every page on this site is informational only. Picks are the model's read of public market data, not a personalised recommendation. Always do your own research and consult a licensed advisor before making investment decisions.

Does the AI know your portfolio?

The Top 15 model does not. It analyses the market in isolation and produces the same list for every user. The separate AI Portfolio Grade feature (Pro) is what looks at your specific holdings and writes personalised feedback.

How is the confidence score calculated?

The model assigns it directly based on how many signals agree and how cleanly they align. It's a heuristic, not a probability — a 90% confidence pick is not 'a 90% chance of being right'. Treat it as the AI's relative ordering, not a literal forecast.

Can the AI hallucinate a ticker or price?

We post-validate every pick: the ticker must exist in the candidate set we sent, prices must round-trip against Polygon, and the JSON shape is enforced. If any check fails, that pick is dropped before you see it.

How is past performance shown?

Every analysis run is snapshotted to a Track Record table. The Track Record tab compares each historical Buy/Strong Buy pick's price-at-pick vs. today's price and benchmarks the average return against the S&P 500 over the same window. There's no survivorship bias — losing picks stay in the data.

We update this page when the methodology changes

Models, prompts, and screening rules evolve. If we change something material to how picks are generated, this page is the source of truth — bookmark it.

See what Pro unlocks

Informational only · Not investment advice · Past performance ≠ future results