Methodology
Most stock-research sites show you their winners. This page is about how the system actually works — sources, scoring, what’s included in the numbers, and what isn’t.
GptInvest started in late 2023 as a question: could a multi-agent AI pipeline pick US equities that consistently beat the S&P 500 over a 6-month horizon? The first Mid-Term portfolio went live on November 9, 2023, with 15 names from the S&P 500. It’s still running, and several other portfolios have been built around the same scoring core since.
Below is how each layer of the system works. If something here is unclear or contradicts what you see elsewhere on the site, that’s a bug — write in.
Where the ideas come from
Each trading day the system scans a few hundred candidates pulled from four parallel feeds:
- News catalysts — earnings releases, FDA decisions, M&A, regulatory actions, guidance changes
- Earnings surprises — beats and misses on revenue and EPS versus consensus, plus the post-release reaction
- Price-and-return anomalies — unusual volume, gap moves, momentum breaks, mean-reversion setups
- Web-search trends — public attention signals via Google Trends and similar
These feeds are noisy on their own. Most of what they surface is junk. The job of the next layer is to filter.
The scoring layer
Candidates are scored by a multi-agent stack across six dimensions: fundamentals, technicals, news and sentiment, earnings quality, macro and industry context, and alternative data. Each dimension is handled by a different model rather than a single LLM doing everything. This isn’t theater — a single model asked to weigh momentum, balance-sheet quality and macro positioning at once produces mush. Splitting the job and recombining the views gives a more stable composite.
The composite score is what drives the rating you see on StockRatings.ai (refreshed weekly) and the daily idea selection on TrendingStocks.info.
How the portfolios are built
Three broad portfolios and three thematic ones. They share the scoring core but differ in what they do with it.
Mid-Term. 15 names from the S&P 500. Position weights are bounded between roughly 2.6% and 10.5%. The target horizon is around six months. Rebalances happen when the conviction stack shifts materially, not on a fixed calendar. Inception November 9, 2023.
Max AI Rating. Holds the highest-rated names in the latest weekly snapshot. Refreshed every Monday. Turnover is higher than Mid-Term by design. Inception July 18, 2025.
Hot Ideas. Equal-weight basket of up to 10 names selected every Monday at 3 PM ET. Heavy on near-term catalysts. Highest turnover of the three, also the highest-variance. Inception June 18, 2025.
Thematic — AI Core, Biotech Growth, Defensive Low-Vol. Same scoring engine, narrower universe. These are not designed to beat the S&P. Biotech Growth, for example, has trailed the S&P since launch — that’s not a failure of the system, it’s biotech being biotech. Their job is targeted exposure with AI-driven selection layered on top.
All portfolios are benchmarked against the S&P 500 (SPY). The broad ones are also compared against the two actively managed AI ETFs that exist as alternatives — Qraft AI-Enhanced US Large Cap and Amplify AI Powered Equity — so you can see whether the system adds anything over the AI-investing alternatives a retail investor could actually buy.
Outcome tracking
Every short-term idea published on TrendingStocks.info has three things stamped on it: an entry date, a tracking window, and — when the window closes — the realised return. Both winners and losers stay visible, with no removals or revisions after the fact. Nothing is quietly deleted when a call ages badly.
This is the part of the system that’s hardest to fake and most useful for anyone deciding whether to trust the output. The outcome log is the trust layer, not the marketing layer, and that’s the right way around.
What’s not included in the numbers
All portfolio returns shown on the site are model and hypothetical. They are calculated from each portfolio’s inception date forward, using closing prices, with dividends reinvested where applicable. They do not reflect actual trading. They do not include:
- Brokerage commissions or transaction fees
- Bid-ask spread or slippage on entries and exits
- Borrowing costs for any short or leveraged positions
- Taxes of any kind, including dividend withholding for non-US holders
- Account-level fees, custody fees, or platform fees
Actual results from attempting to replicate any portfolio would differ, and on high-turnover strategies the difference can be material. The headline excess return is the upper bound of what’s achievable, not the realistic one. Read the numbers with that in mind.
Backtests versus live performance
Everything on the portfolio scoreboard is live, time-stamped from each portfolio’s actual inception date. The only backtested material on the site is in the Quant Lab section, and it is labeled as such. Live and backtested performance are never combined into a single headline number — that’s a hard line.
Backtested results carry inherent limitations: the benefit of hindsight in strategy construction, the inability to fully account for the impact of trading on historical prices, the use of universes that may exhibit survivorship bias, and assumptions about execution that may not hold in live markets.
What the AI does and does not do
The scoring engine combines outputs from multiple language models and quantitative filters into a single composite. It identifies patterns in historical and current market data. It does not predict the future, does not have access to private or non-public information, does not have privileged insight into corporate actions, and is not capable of judgment about whether a position is suitable for any specific person. Where the system is wrong, it is wrong silently — it does not flag its own uncertainty.
The phrase “AI rating” on this site refers to a quantitative score, not a recommendation to buy or sell anything.
What this is not
It’s not personal investment advice. Nothing on the site is calibrated to your situation, risk tolerance, time horizon, tax setup, or income needs. The ratings are quantitative outputs of a system, not recommendations.
It’s not a guarantee. Mid-Term has beaten the S&P 500 since November 2023; that does not mean it will continue to. AI systems are pattern recognisers, not oracles. Anyone telling you otherwise — including any future version of this site — is selling.
It’s not a black box. The sources, scoring dimensions, rebalancing logic, and benchmarks are documented on this page. If a number or rule isn’t explained here or on the relevant sub-domain, that’s an oversight. Email and it’ll be fixed.
Regulatory status
GptInvest is a research publisher. It is not a registered investment adviser, broker-dealer, or financial planner in any jurisdiction. Materials are published on a regular basis to a general audience and are not tailored to any individual. If you need advice calibrated to your personal situation, talk to a registered investment adviser, broker, or qualified financial planner in your country of residence.
Investing in securities involves the risk of loss, including the loss of principal. Concentrated portfolios — including all portfolios on this site, which hold between 10 and 15 positions — carry materially higher single-name risk than broadly diversified index funds.
The full set of disclaimers, including jurisdictional limits and limitations of liability, is in the Disclaimer and Terms of Service.