← Blog · 2026-07-19 · 7min
How biweek's AI Picks US Stocks — Walk-Forward Validation, Honestly Explained
- Two deep-learning models rank US stocks daily by modeled probability of +10% within 14 trading days — Steady (8 picks) and Aggressive (10 picks).
- We only trust walk-forward validation: train on the past, test on unseen future periods, retrain per fold. Three consecutive simulated years above the S&P 500.
- Honesty first: hit rate is ~4 in 10, the edge is asymmetry (−7% hard stop, 12% trailing stop), and all performance figures are simulations — every limitation is disclosed below.
There's no shortage of "AI stock picks" on the internet. Most of them share a problem: you can't tell how they were tested, what their hit rate is, or what happens when they're wrong. This post explains exactly how biweek works — including the parts that don't flatter us.
1. What the models actually predict
Every trading day, we scan thousands of liquid US-listed stocks (minimum 140 trading days of price history — no fresh IPOs) and estimate, for each one, the probability that it rises +10% within the next 14 trading days. That's the entire prediction target. No price targets, no "this stock will change the world" narratives — one falsifiable question, answered with a probability.
Two models answer it independently: Steady (trained toward consistent, lower-drawdown names — 8 picks) and Aggressive (trained toward bigger movers — 10 picks). They share exit rules but select different stocks.
2. Why walk-forward validation is the only test we trust
Anyone can fit a model that looks brilliant on the past. The honest test is: train on data up to a date, then trade only the unseen future — and repeat, retraining each period (a "fold"). That's walk-forward validation. Our models were tested this way across three consecutive years, and in simulation each year finished above the S&P 500. We re-validate continuously as new data arrives.
3. Missing 6 of 10 calls — and still making money in simulation
The models' hit rate is roughly 4 in 10. That sounds bad until you see the exit rules: every position has a −7% hard stop, a 12% trailing stop from its peak, and a 60-trading-day cap — whichever comes first. Losers are cut small; winners are allowed to run to +20%, +40%, occasionally +100%+. A few big winners can cover many small losses. This asymmetry only works if you hold the entire list — buying one or two picks breaks the math, which is why we always show the full ranked set.
4. What we disclose vs. what we protect
| Disclosed | Prediction target, validation method, exit rules, simulated performance and drawdowns, every known limitation |
| Protected | Input features, network architecture, training details — the competitive core |
We publish the what and the proof, not the how. If we published the how, anyone could clone it; if we hid the proof, you shouldn't trust us.
5. Known limitations (the honest list)
| Simulation gap | Idealized fills — real execution is worse |
| Regime risk | Patterns learned from history can stop working |
| Drawdowns | Simulated max drawdowns −21% (Steady) / −26% (Aggressive); near −40% in bear markets |
| Not advice | Probabilities inform; they don't decide. Your capital, your decisions |
6. Try it
Today's ranked picks are public — no login, updated every trading day: biweek.app/en/candidates. Member features are invite-only for now; the waitlist is on that page. Predictions, not guarantees — but honest ones.
FAQ
Curious what the deep-learning models picked today?