Do AI stock signals actually work?
Short answer: yes, but only modestly, and only when they are filtered honestly. Cashora's signal engine was directionally correct 58.8% of the time across 2,289 positions on 30 US stocks and ETFs over ten years of daily data, averaging +1.52% per 20-trading-day position at a mean Sharpe ratio of 0.59. On the 15 tickers deliberately held back from every tuning decision, the hit rate was 57.3%. Unfiltered, the same engine measured roughly 51% — indistinguishable from a coin flip.
Last measured 2026-09-22 · Thomas Solutions LLC
What makes the difference between 51% and 58.8%?
Three filters, each validated on data that was not used to choose them. Long calls only — short calls measured about 45% accuracy and are never marked actionable. The price must sit above its 200-day average, which removes most losing trades taken against the broad tape. And the confidence score must be 60 or higher. Removing signals turned out to matter far more than adding indicators.
Can AI predict stock prices?
No. Nothing predicts a price. What a measurable engine can do is estimate the probability that the next few weeks lean up rather than down, and be honest about how often that estimate has been right before. That is why every Cashora score is a probability with a grade attached, not a forecast or a target price.
Is a 58.8% hit rate good?
It is realistic. Published professional edges on short-horizon directional calls sit roughly between 53% and 58%. Anything advertising 80–90% accuracy on daily or weekly direction is either overfitted to its own test data or measuring something much easier than a real trade. We explain the arithmetic in what counts as a good hit rate.
Where does the model lose money?
CVX, T and HD lost money per position out of sample, and INTC and T finished below a coin flip. On most large-cap names, simply buying and holding over the same decade beat the signal portfolio — NVDA most dramatically (+9,542% held versus +5,520% traded). The full per-ticker table, winners and losers, is published on the measured performance page.
How were these numbers produced?
Walk-forward replay on daily split-adjusted Yahoo Finance closes from 2016-09-19 to 2026-09-18, 2,514 bars per ticker. On each historical bar the engine sees only data up to that bar. A position enters at that day's close and exits at the close 20 trading days later, one fully-allocated position at a time so overlapping trades cannot inflate compounded returns. You can reproduce any of it on the Backtesting page.
What should someone realistically expect?
A small statistical tilt, visible over dozens of positions and invisible over three. Drawdowns of 20–50% appeared in the measured history of most tickers. Signals are a timing research tool, not a substitute for owning good businesses.
Hypothetical back-tested results, excluding commissions, slippage, taxes and dividends. No Cashora user achieved them and past performance does not indicate future results. Cashora is a market probability signal predictor operated by Thomas Solutions LLC — not a broker, no trades are executed, and nothing here is investment advice. See our Terms of Service.