The close lands inside the 7,604–7,745 range the model published that morning: the S&P 500 falls 0.48%

Validated over 1073 out-of-sample sessions, walk-forward: each day is predicted training only on earlier ones, retraining every 21 sessions. It is right 59.83% of the time against 51.16% for the majority class (always calling the more frequent outcome). Brier 0.23457 against 0.24986 with no model.

Of the 139 variables computed (the index, the 500 stocks one by one, VIX, rates, gold, copper and the calendar) the model keeps the 120 most correlated with the target every 21 sessions, chosen using past data only. Today's are listed one by one in prediccion.json.

The ranking is recomputed every session, and this is what moved from one to the next. Next to each move is the factor that produced it: momentum changes every day, value whenever the price does, and quality and growth only when new accounts come in. Why the price moved is not in this data, so it is not told.

The biggest moves in the whole index: DDOG climbs 36 places, from #243 to #207 (Momentum 5.7 → 6.1) and COHR drops 69 places, from #109 to #178 (Momentum 8.9 → 8.0).

Source: rankings_diarios.json.gz · rank by multi-factor (the average of the four), across the 465 companies with all four computed · scores on the 0-10 scale used on company pages, where 5 is the index average. See all 500 scored →

The section above shows the ranking moving; almost always it is the price moving it. This is the other thing: a quarterly or annual report rewrites value, quality and growth at once, and those are audited accounts, not a quote. Here is the before and after, on the 0-10 scale used on company pages, where 5 is the index average.

Filing date on EDGAR, not the quarter end. Scores come from the rankings calendar: the same ones each company page publishes. No opinion on the results, only the measured effect.

The rare thing on a Form 4 is a purchase. The largest on September 4, 2026: Cascade Investment, L.L.C., Gates William H Iii, 10% Owner at Republic Services, bought 365,039 shares at $223.16 (81.46 million dollars), code P, meaning on the open market and with their own money. see the Form 4 →

Not a one-off: a campaign. That same buyer has bought Republic Services on 20 consecutive trading sessions, Aug 10, 2026 to Sep 4, 2026: 4,952,848 shares for 1,091.4 million dollars, and after the last one declares 115,756,830 shares held. It is the largest by money of the 11 runs of three or more consecutive sessions in the Form 4 filings downloaded between Aug 25, 2025 and Sep 8, 2026. That someone buys says nothing about where the price goes: what is checkable is that they filed it with the SEC session after session.

4 purchases were filed that day across 4 companies. Over the 23 days the file covers, Aug 17, 2026 to Sep 8, 2026, there are 189 purchases worth 1,169 million dollars across 64 companies, against 1984 sales: which is why a purchase gets told and a sale gets tabulated.

The biggest sale of the day tells you nothing. Shek Bernard, of Sandisk, sold 4.1 million dollars of stock and still holds 25,305 shares: that is 8.36% of the position, and the sale was scheduled months in advance under a 10b5-1 plan. That is tax plumbing, not a view on the company.

The only sale filed that day was scheduled.

Source: Form 4 filings with the SEC · alertas_insiders.json, session of Sep 8, 2026. Every name links to its filing on EDGAR: the claim and the proof on the same line.

The context over those same 23 days: 1984 sales worth 18,016 million dollars (3 with no declared price, outside that total), of which 792 were discretionary, and of those 208 exceeded 25% of the seller's position. That is where to look.

All five beat the benchmark (+467.6%) on cumulative return over fifteen years, but cumulative return is not the proof: the benchmark carries the bias explained below. The weakest is Value 100, with an alpha of -1.28%: negative, meaning what it earns does not pay for the market risk it takes to earn it. Publishing the figure that comes out worst is the only reason the others deserve any credit.

Source: carteras.json · index based at 100 on Mar 4, 2011. Walk-forward with point-in-time data: on each date scoring uses only prices up to that day and fundamentals already filed with the SEC. 185 rebalances through Aug 17, 2026.

Over the past month the value portfolio lost -1.79% while momentum gained +1.74%. Across that six-month window the benchmark rose 9.7%, so this is not the whole market moving: it is a rotation inside one.

The extreme of the series is July 9, 2020, six years ago, at +30.2 points. And what happened after each extreme? Counted, not assumed: of the 4 extremes in momentum's favor with enough data after them, in 3 the lead had flipped sign six months later; of the 5 extremes in value's favor, in 4. That is a count over 9 episodes, not a rule.

Computed from carteras.json · 185 monthly rebalances between Apr 4, 2011 and Aug 17, 2026 (fifteen years) · 179 overlapping windows of 6 rebalances · recent stretch: the last 24 rebalances, the shortest run that still holds 4 non-overlapping windows.

With the rank and the score they carry before reporting. Section 03 will publish the after once the document reaches EDGAR, which is zero to three days later: the press release does not move our scores, the filed accounts do.

Source: calendario_resultados.json · Nasdaq public earnings calendar. Earnings date announced by each company, not the SEC filing date, which is not published in advance. Scores from the Sep 9, 2026 ranking.

• What insiders did yesterday. Form 4 can be filed up to two business days after the trade, so the freshest data is always the day before last.
• Whether momentum keeps winning. Fifteen years and 185 rebalances are not enough to claim a factor mean-reverts. We can describe what happened, not what comes.
• What any company is worth. This report computes no price targets. No Monte Carlo, no DCF, no fair value. When you see one elsewhere, ask which cost of capital discounted it and who chose that number.
• Whether the S&P rises tomorrow. We measured it with the same model (139 variables) and the same walk-forward protocol, over 1,073 out-of-sample sessions: the model calls direction right 53.77% of the time and saying "up" every single day is right 53.77%. The gap is noise: there is no signal, so we don't sell one. Both figures are recomputed every edition.