The claim sounds like a gimmick: an AI that predicts stock prices “to the penny” 21 trading days out. That’s a month of market chaos — Fed meetings, earnings surprises, tariff tantrums, black swans — all compressed into a number with two decimal places. The product making the claim is called Predictive Alpha, and it comes out of Keith Kaplan’s TradeSmith.

But Keith Kaplan’s TradeSmith has been building this thing for 26 years, backed by tens of millions of dollars and tens of thousands of man-hours, with a team of 69 engineers, 22 ML/AI specialists, and 23 quants — former Wall Street traders, PhDs, a retired lieutenant colonel who ran top-secret nuclear missions, a molecular geneticist, an eye surgeon.

The AI is called An-E — short for Analytical Engine. The product Kaplan sells around it is called Predictive Alpha. It’s trained on 1.3 quadrillion data points and more than 50,000 backtests. It covers over 2,300 stocks, funds, and ETFs. And Kaplan’s pitch is that the old buy-and-hold playbook is done. For the companion model write-up, see An-E AI: TradeSmith’s 21-Day Price Forecasts, and for the product-level verdict, see the Is TradeSmith legit review. For the dedicated product review, see Predictive Alpha Review: Keith Kaplan’s AI Tool.

He might be right about the timing.

The Lost Decade Thesis

In November 2025, Goldman Sachs strategist Peter Oppenheimer — who correctly called U.S. underperformance in 2023 — published a bleak forecast: the S&P 500 would return roughly 6.5% annualized over the next decade. That’s barely a third of the long-term average. Goldman’s earlier forecast from David Kostin in October 2024 was even grimmer — just 3% annualized. Bank of America piled on in December 2025, projecting the S&P 500 would shed 0.1% over the next ten years.

The logic is straightforward. The S&P 500’s CAPE ratio sits near the 97th percentile of all historical readings. Market concentration is extreme — the Magnificent Seven have driven the bulk of returns, and history shows that when leadership gets this narrow, the following decade tends to be a washout. The pattern repeats: the 1970s followed the Nifty Fifty, the 2000s followed the dot-com boom, and the outcome was boringly consistent.

If Goldman and BofA are right — and they have been wrong before, but the math is harder to argue with than the forecast — then the next decade will be brutal for anyone who just buys and holds an index.

Which is exactly where Predictive Alpha enters the picture.

What Does “Predicting to the Penny” Actually Mean?

The “to the penny” framing is the headline version of the claim. The fine print describes something more specific.

TradeSmith’s own materials break it down. Predictive Alpha generates a “Prime Projection Date” — the specific day within the 21-day window where the model has the highest confidence. It gives you a price target, a directional accuracy percentage, and a historical target accuracy percentage. The overall average directional accuracy across all rated stocks is about 57.5%. The average historical target accuracy — how often the stock actually touches its projected price within the window — is about 60%. Past performance does not guarantee future results, but those are real numbers, and a 60% hit rate on price targets, consistently, over thousands of stocks, across multiple market regimes, would be genuinely useful. The system uses an ensemble of two AI models — one focused on longer-term trends, one on day-to-day volatility — working together to produce a single forecast.

The 85% figure that appears in some of the marketing is backtested, on a specific subset of picks, with a specific strategy — a different claim than “the model is 85% accurate on all stocks.” The product documentation breaks the figures down separately. Past performance does not guarantee future results.

How Real Is This Compared to the Rest of the Field?

Academic research on AI stock prediction has come a long way. Lopez-Lira and Tang (2023) showed that ChatGPT-4 can predict next-day stock returns from news headlines with a Sharpe ratio above 3. MarketSenseAI, a GPT-4 framework tested on S&P 100 stocks, delivered 10-30% excess alpha. StockGPT, a transformer trained on 70 million daily stock returns, produced significant alphas even 23 years out of sample.

But there’s a catch. The Chronos paper (2024) showed that while LLMs detect short-term reversal patterns, the edge disappears after accounting for just 3 basis points of trading costs. The gross Sharpe of 3.17 became a net -1.49. The market is brutally efficient at the micro level.

TradeSmith’s An-E sits in an interesting middle ground. Rather than predicting tomorrow’s open from a headline, it forecasts the shape of a stock’s price trajectory over a month — a timeframe where momentum and mean reversion operate, and tick-level noise smooths out.

The Demo Picks in Context

The promo materials highlight specific calls: Applied Digital was called for +16% and delivered, SoFi was called for +9% and hit, and Carvana was called for +25% over two days. Past performance does not guarantee future results.

I checked the data. SoFi stock skyrocketed 47% in November 2025 alone, hitting an all-time high. Carvana was up 65.7% year-to-date. Applied Digital was up over 300% for the year. These were stocks in strong uptrends, riding massive tailwinds. An AI calling them higher is impressive, but it’s not divination. The model identified momentum early, and momentum continued.

The more interesting test would be the bearish calls — the stocks An-E flags to avoid. That’s where the real value would be, especially with Goldman predicting a lost decade.

The 1.3 Quadrillion Data Points Claim

If you’re tracking tick-level data — every trade, every quote, every spread change — for 2,300 stocks over 26 years, you accumulate an astronomical dataset. The S&P 500 alone generates millions of trades per day. Multiply by thousands of stocks and decades, and you’re in that territory.

The real question is whether the model uses all of them well. Deep learning systems can ingest massive datasets and extract real signal. The bigger risk is overfitting on noise. That’s why the 50,000+ backtests matter — they stress-test An-E across bull markets, bear markets, crashes, and recovery periods to make sure it’s learning actual patterns, not artifacts.

What the Demo Tells You

Keith Kaplan has built something real. An-E is not a gimmick, even if the marketing sometimes oversells it. The team behind it is serious — 69 engineers, 22 ML experts, 23 quants, plus a cast of domain experts that reads like a Tom Clancy novel. The 26 years of development and 1.3 quadrillion data points represent a genuine R&D investment.

The product acknowledges its own limitations, with historical target accuracy averaging 60% and directional accuracy averaging 57% — probabilities rather than magic numbers. In a world where the S&P 500 might return 6.5% annualized for the next decade, a system that gives you a 60% edge on individual stock moves over a 21-day window is worth paying attention to.

The question is whether a 60% hit rate on monthly price targets, applied consistently across thousands of stocks, can outperform a buy-and-hold strategy that Goldman Sachs says is heading into a lost decade.

The question of whether a 60% target-accuracy model carries a usable edge over a buy-and-hold baseline into a Goldman-projected 6.5%-return decade is the one the backtest cannot answer on its own.

The Predictive Alpha demo opens the forecast history to inspection — the 21-day projections, the directional and target accuracy averages, the stocks the model tracks. The thesis that An-E carries an edge over a buy-and-hold baseline into a 6.5%-return decade rests on whether that 60% target accuracy holds consistently across market regimes the backtest did not cover. For the full index, see Promo Watch.