Proof
Open the app

How it works

What the model does, how it was tested, and — the part most tools leave out — where it is weak.

Educational research tool — not investment advice, and not a recommendation to buy or sell any security. Backtested results are hypothetical, come from one 18-year dataset the model was also tuned on, and are not a promise about the future. The forward track is a model portfolio and a sandbox paper account — no money is invested in either, so no figure here is a return anyone received — and it is short. You can lose money. Do your own research.

The short version

After each market close the ~800 most liquid US common stocks are scored on the model's weighted themes — value, quality (growth, for younger companies), momentum, size, capital discipline, institutional positioning and insider activity, each carrying an equal share. Each theme is built from individual numbers, each number is standardised across the names scored that day, and the themes are combined into one composite. The Hot Stocks tab is that ranking. Two of the seven — institutional and insider — have no live data source yet, so today's live ranking runs on the other five while the backtest used all seven; the Hot Stocks tab flags this on every scan.

The research also constructs a book from that ranking — the top decile of the large-cap tier, score-weighted, capped at 8% per name — and tracks it forward against the S&P 500. Both are shown in the app: the Valquo Index tab lists the book and the Track Record tab its forward record. They are a model portfolio with no money in it, published so the method can be checked, not as a recommendation to hold any of it.

Point-in-time, and why it matters more than the score

The backtest is built from a point-in-time fundamentals panel: on any historical date the model sees only what was actually public and filed by then. Restated figures do not leak backwards. Institutional (13F) holdings are lagged by their real filing deadline — the panel's effective lag is about 111 days, more conservative than the 45-day rule. We checked this the direct way: feeding the model fresher, not-yet-filed holdings makes it weaker, not stronger, which is the opposite of what a look-ahead artefact does.

Survivorship

Delisted and acquired companies stay in the universe up to the date they disappear, using a delisting mask rather than today's surviving-company list. A backtest run on the names that still exist is a backtest of not going bankrupt. The live forward track has the same property by construction: the picks are dated when they are made and measured forward, so nothing can be quietly dropped after the fact.

Costs — quote the breakeven, not the net number

Trading costs are modelled, not assumed away. The top decile's breakeven cost is about 134 basis points one-way against a measured 33 bps real-world cost profile, on roughly 261% annual turnover. We lead with the breakeven deliberately: it requires you to believe no particular cost estimate. Borrow cost is not modelled, which affects the long/short statistic but not the long-only book that the Index actually is.

How we decide something is real

The one result that had to clear a threshold written down in advance

The composite is roughly one-seventh each of value, quality, momentum, size, investment discipline and institutional ownership — nearly the standard factor set — so "it beat the average" could not distinguish a real edge from an assembly of known premia. Two write-ups were prepared before the regression ran: one to publish if the intercept cleared Newey–West t > 2, and one saying the honest description was efficient factor exposure if it did not.

It cleared it. Against the Fama–French five factors plus momentum, the top-decile spread carries an intercept of +6.99%/yr (t = 3.98) over 68 non-overlapping windows, 2009–2025, with all six pre-registered specifications positive at t > 2 and spanning +5.1% to +10.9%. Run through the same pipeline, a passive large-cap ETF returns +0.68%/yr (t 1.58) — the placebo that shows the machinery reports nothing significant when there is nothing there.

Read that as a research finding and nothing else. It is measured on a historical simulation, not an account. It is not an expected return, not an achievable return, and not a return anyone earned. It is one panel; the t-statistic is not corrected for the equity trials logged behind it (248 by 2026-09-29, and the count keeps rising — the live figure is on the Proof page); the panel's known-contaminated early period has been removed from the sample outright rather than discounted, and the conservative single figure is the first half's +5.19% if you want one number. And it does not mean the strategy beats the factor ETFs you can actually buy — that was tested separately and was not demonstrated (a +9.2pp margin at t 1.10, negative in the first half: a null).

How long the edge lasted, and what that does not license

Every headline figure this project publishes is measured over a 63-trading-day forward window, which was an inherited default rather than a measured optimum. So the composite was scored at 8 horizons — one quarter through two years — from a single panel build, on the set of dates observable at every horizon, so that the length of the window is the only thing that varies. The result: CONSTANT-RATE. In the backtest, the top decile of the hot list beat the equal-weighted universe by about 6.6% annualized over the next three months — and was still ahead by about 5.1% annualized two years later — even though a given name typically stays in the top decile for only one quarterly rebalance.

An independent route reaches the same place without touching the decile machinery: the median per-date rank correlation between score and forward return rises with horizon, 0.0336 at one quarter to 0.0655 at two years. And the alpha is well measured at every horizon — its t-statistic never falls below 3.16, and is 3.83 at two years — so this is not a signal that survives because its error bars widened.

Three limits, and the third is the one people get wrong. Measured on the corrected 2,531-name / 69-date panel: long-only top decile versus the equal-weighted universe; gross of costs; and it is the same single in-sample panel every other published figure comes from — not a forward test. Second, a longer forward window is not new data: the eight horizons are eight views of one sample, not eight samples, and the windows overlap almost entirely. Third: It is not a finding that the book should rebalance less often — the list is re-ranked every quarter, and what was measured is how one quarter's selection went on to do, not a comparison of holding policies. What was measured is the buy-and-hold return of the names picked on a single date; a quarterly-rebalanced list re-picks and compounds fresh selections, and those are different claims. Only the first was tested.

The options side wins about a third of the time, and that is the design

The research also runs an options book, and its hit rate is 35-37% depending on which book you measure. That number is the one most likely to be misread, so here is the whole distribution rather than the average. Over 3,885 simulated trades on 187 names, 2016-01 to 2025-10:

The middle trade loses 52% of the premium. The trades that at least doubled are 87% of everything the winners made. A book like this is supposed to lose most of the time, and a string of losses is its normal texture rather than a sign it has stopped working.

How long is an ordinary bad run? Expect losing streaks. Over 20 trades the typical worst run is 5 in a row, 44% of stretches contain a run of 6 or worse, and the record's worst at this scale is 20. Losing runs are longer than a coin-flip model predicts, because trades opened near each other in time share a market: the clustering measures 2.667 against a shuffled null whose 95th percentile is 1.244 (1,000 shuffles, p < 0.001). Assuming independence would put the 95th-percentile worst run at 10 instead of the measured 12 — so the tidy arithmetic is the one that would cry wolf.

None of that says the options alerts work — they were tested and they do not. Measured against random entry on the same names and dates, the alert's choice of day subtracted value: −5.06 percentage points per trade, paired sign test p < 0.00001. We publish the payoff shape so a losing streak is legible, not as evidence of an edge. There is none demonstrated here, and the alerts are an idea generator rather than a signal to act on.
Streaks measured on the 6,032-trade corrected-era random-entry control (the real book's own per-trade sequence is not banked); the control hits 37.2% against the book's 35.3%, so these runs are if anything too SHORT.

Where it is weak — read this part

What this is not

It is not a forecast, not personalised advice, and not a claim that the next twelve months will resemble the backtest. It is a disciplined, auditable ranking with its assumptions and its failures written down. If a page here ever shows you a number without telling you whether it is backtested or live, that is a bug — please report it.

It is also not a signal service. The app does show a model portfolio, options alerts and forward records — but every one of them is a paper record: a model book priced at closing marks, and a broker sandbox account with no real money. They are published so the method can be checked against what actually happened next, and they are labelled thin until they are long enough to mean anything. None of it is a return anyone received, and the options alerts in particular were tested and found to lose to random entry (above). Treat everything here as analysis you can check, not something to act on.

← Back to the app