Look up any US stock: performance over every horizon, how it compares with the market and its sector, and an AI breakdown of why it moved.
Breakdowns written today and yesterday. Click one to open it. ✓ means every figure in the text matched the data.
Look up any US-listed ETF, including the ETF versions of the big index funds (VOO, VTI, VXUS).
Breakdowns written today and yesterday. Click one to open it. ✓ means every figure in the text matched the data.
Daily closes. VOO and VTI are Vanguard's ETF versions of its index funds.
Look up a major coin: performance over every horizon, compared with Bitcoin, with an AI breakdown of the move.
Breakdowns written today and yesterday. Click one to open it. ✓ means every figure in the text matched the data.
Top 100 coins by market cap, 24h change.
The AI forecasts each business five years out; code ranks them. Top 5 held.
Every quarter, after earnings season, the AI reads each company's current fundamentals and recent news and forecasts two things: how fast revenue will grow over the next five years (with a low, middle, and high estimate) and what its profit margin will be in year five. It never says "buy" or "sell."
Code then values each company the same way: today's revenue grown at the middle estimate, times the forecast margin, at one uniform price-to-earnings multiple, compared with today's market value. That gives an implied yearly return, used only to rank. The five highest-ranked stocks are held, equally weighted, until the next quarter.
The idea comes from FutureSearch's study of the S&P 500, whose central lesson is that short-term moves are close to unpredictable while long-run business forecasts may not be. Results here will take a year or more to mean anything.
| Rank | Stock | Revenue growth / yr (low, mid, high) | Year-5 margin | Implied return / yr |
|---|
The paper-trading experiment that came before the research tools, kept in full as a case study.
From Aug 18 to Oct 20, 2026, the AI made daily market calls on a fixed watchlist and placed simulated trades on them (Alpaca paper account, no real money), graded every night against real prices and against simply holding SPY.
Result: over … closed stock trades it lost …, while the same dollars in SPY made …. On the day it was stopped (159 trades), its per-trade return versus SPY was below zero with 95% confidence, and a mid-test rule change hadn't fixed it. The AI also graded its own calls more generously than prices did (… vs … price-checked). The daily strategy ended on Oct 20, 2026; the complete record is below, unedited.
What came out of it: daily price moves look close to unpredictable from prices and news, which is why the project now focuses on research and on the longer-horizon experiment.
Each trade vs. SPY over the same window. Market-adjusted removes the market's own move.
The AI's grade of its own calls vs. what prices actually did.
| The AI's confidence | Calls | Direction right | Avg trade return |
|---|
Every change, dated, so results can be read against it.
Click a header to sort. Small samples are flagged.
Tickers with 3+ trades.
Each pick gets a confidence level and an event-risk flag from the AI. Size is cut for low confidence, event risk and shorts, because each lined up with worse results in this system's own history.
Loading…
No sizing rule changes until both sides of a comparison have at least 20 closed trades. Smaller groups are labeled small sample.
Every request and its outcome. A pass means nothing cleared the bar.
Sell closes at market; after hours it fills at the next open. PIN required.
Market value vs. cost basis, updated nightly.
Every request and its outcome. A pass means nothing cleared the bar.
Every night, 10 coins, long-only, $500 per pick. Tracked separately from stocks.
Cumulative P&L, nightly and on-demand trades combined.
Nightly and on-demand trades.
Every request and its outcome. A pass means nothing cleared the bar.
What Great Lakes Investments is, how it works, and what it isn't.
Great Lakes Investments is a student-built AI market research project. Type in a stock, ETF or major coin and it shows how it has performed over 1 day, 10 days, 30 days, year to date, 1 year and 5 years, how that compares with the market and its sector, and a written breakdown of why it moved.
It started as an automated paper-trading experiment. That experiment is finished and kept in full under Track Record, including the result: it didn't beat the market.
1. Prices: up to five years of daily prices come from Alpaca's market data, adjusted for splits and dividends.
2. Numbers: code, not AI, computes every return, the comparisons with SPY and the sector, volatility, drawdown and the 52-week range.
3. News: the last ten days of company headlines and the latest market headlines come from Finnhub.
4. Breakdown: An AI language model reads those numbers and headlines and writes the explanation. It is told to use only the numbers it's given, to cite the news it relies on, and to say plainly when nothing in the news explains a move.
Numbers come from data, not the AI. The performance tables are calculated directly from prices. The written analysis can be wrong; the numbers it sits next to are not invented.
Every figure in the AI text is checked automatically. Code pulls each percentage, dollar amount and multiple out of the breakdown and matches it against the computed data and the news the AI was given. Reports are badged "all figures match" or list exactly which figures couldn't be matched. This came directly from the paper-trading experiment, where the AI graded itself more generously than prices did.
"No clear catalyst" is an allowed answer. Every breakdown is labeled with how strong its explanation is.
No recommendations. It explains what happened. It never says buy, sell or hold.
The failed experiment stays public. The Track Record tab shows the paper-trading loss and how the AI graded itself more generously than prices did.
Fund lookups cover ETFs; mutual fund share classes such as VFIAX aren't in this data source. Crypto covers about 20 major coins, and its history may be shorter than five years. Prices can be delayed by 15 minutes or more during the trading day. To keep costs bounded, each visitor can generate a few new breakdowns a day, and every breakdown is shared and reused for the rest of that day.
An AI language model (via the Anthropic API) for analysis, Alpaca for market data, Finnhub for news and company data, a Cloudflare Worker for lookups, GitHub Actions for the scheduled long-horizon track, and GitHub Pages for this site. The code, a candid status memo and the roadmap are public on GitHub.
Enter your PIN to run this on-demand analysis.