Buy at the close.
Sell at the open.
Every trading day the model ranks the S&P 500 and picks the ten strongest stocks, bought in the closing auction and sold at the next morning's open. Two brakes decide how much of the account goes in. Below: tonight's list, a paper account that tracks every pick from its first day, and what taking every trade since 2017 would have done, costs included.
The honest answer first
Loading…
Starts with the first list published before its order deadline. Nothing is backfilled.
The alternative balances are shadow comparisons, not recommendations. On the S&P 100, rank weighting returned 15.3% historically with the brakes and 19.6% without them; the latter also suffered a 31.3% drawdown. Both were discovered in those same data, so only their forward records count now.
Tonight's picks in the news
The latest headlines from Yahoo Finance's news feed for each pick; some are general market stories that mention it. The picks don't use news. Every stock's headlines are being saved each day, so that months from now news can be tested on data nobody has looked at, which is the only honest way to test it.
Challenger: machine-learning picks
Paper only · not the main ruleA random forest trained on 22 signals (momentum at several horizons, this morning's gap, volatility, volume surges, distance from the 52-week high, market state) picks 10 stocks each trading day, bought at the close and sold at the next open, always fully invested. It needs this morning's opening prices, so its list appears after 9:35 AM ET.
Always invested: SPY by day, the picks by night
Paper only · strict conditionsThis one uses the previous rule's S&P 100 list, because its test was registered before the S&P 500 version existed. The money never sits in cash: hold SPY during the trading day, swap into tonight's picks at the close, and back into SPY at the open. It is the only idea here that beat simply holding SPY by more than luck in the old data, and only when trading costs are close to zero.
| Cost per trade | Per year, 2009–2026 | Holding SPY | Average month vs holding (plausible range) |
|---|
- Four trades a day. At 0.01% per trade it only ties holding SPY; at 0.02% it loses badly. It needs a broker with closing- and opening-auction orders and close to zero commission.
- It's a day trade every day. Buying and selling SPY in the same session counts as day trading. Under FINRA's pattern-day-trader rule that needs a margin account with at least $25,000; check whether that rule has changed.
- The idea was found after looking at all the old data. It passed in both halves (2009–2017 and 2017–2026), and the gain traces to the picks themselves. But only its forward record, starting with the night of October 1, 2026, is untouched evidence.
- Taxes. Every gain is short-term. None of that is in these numbers.
If you had taken every trade
Historical simulationLoading the simulation…
| $10,000 start | Ending balance | Per year | Worst drop | Return per unit of risk |
|---|
Does it survive trading costs?
| Cost per trade, each way | $10,000 became | Per year | Worst drop |
|---|
It trades every night, so costs are paid about 500 times a year. The headline uses 0.01% each way. Auction orders cross no spread, but a broker's commission can be much more than that on a small account.
Year by year
| Year | Picks | SPY hold | Avg invested |
|---|
Each year is its own $10,000 start. The latest year is partial.
Chosen first, checked after
| Period | Nights | Picks per year | SPY per year | Worst drop | SPY worst drop | Avg invested |
|---|
What else was tested
All on the S&P 100 as it actually stood each year (except the first row, which uses the S&P 500), 2009 to September 2026, $10,000 start, costs included. Tested September 30, 2026. Only one beat simply holding SPY by more than luck, and only at near-zero trading costs.
| Approach | Per year | Worst drop | Verdict |
|---|---|---|---|
| SPY buy & hold | +14.8% | −33.7% | The benchmark |
| This page's rule: the same picks from the S&P 500 (top 10, both brakes) | +15.1% | −27.1% | About what SPY made, with smaller drops. Tested October 10, 2026 on S&P 500 stocks no rule had been tried on |
| Previous rule: S&P 100 only (top tenth, both brakes) | +12.5% | −14.8% | Less money, much smoother |
| Same overnight picks, no brakes | +15.7% | −30.4% | Slightly more money, SPY-sized drops, and costs hurt it more |
| Hold the top tenth by momentum, rebalance monthly | +16.8% | −29.3% | Most money of anything tested, but it lost to SPY in 2009–2017 and fell far behind in the spring-2009 rebound. Its lead (+0.17% a month, plausible range −0.22% to +0.58%) could be luck |
| Hold the top fifth by momentum, rebalance monthly | +14.1% | −29.4% | No better than SPY |
| SPY only when above its 200-day average | +9.2%* | −20.7%* | Less money; whipsaws in and out |
| SPY by day, these picks by night (always invested) | +21.8% / +15.8% | −36% | At 0.005% per trade it beat SPY in both halves; at 0.01% a tie; at 0.02% it loses. Tracked on paper above |
| Machine learning: 4 model types × 2 trade types, 22 signals (2012–2026) | +18.1% best | −31.9% | Best of 8 beat SPY's +15.0%, but within luck; every full-day version lost to SPY with ~50% drops. Running as a paper challenger above |
| Wider lists, other trend measures, skipping earnings nights | — | — | None beat the simple rule |
* 2017–2026 only. Monthly portfolios assume 0.10% per round trip on what is traded.
Live paper record
DownloadOne row per night, graded at the next open (provisional) and again after that day's close (final). Fills are the auction prices; 0.01% each way is taken off what was invested.
| Night | Invested | Bought | Result | P&L | Balance | SPY overnight | Status |
|---|---|---|---|---|---|---|---|
| No graded nights yet. The first result arrives after the next open. | |||||||
Tonight's full ranking
| Rank | Stock | Sector | 12-month trend | Last close | Tonight |
|---|
Method, costs, and limits
Prices: Yahoo Finance daily bars (split-adjusted, not dividend-adjusted). Index membership: the Wikipedia S&P 500 (and S&P 100) component tables as they stood each January. No orders are placed. This is a research record, not investment advice, and PontisAi is not a financial adviser. Past results, simulated or live, do not predict future returns.