How the screen works, and what testing it showed
Triple Point flags a stock only when three independent methods agree: CAPE for value, the Piotroski F-Score for quality, and Point & Figure for timing. Any leg that is neutral, dissenting, or uncomputable yields Hold. There is no weighting and no averaging — two legs cannot outvote a third.
CAPE — value
Price divided by the mean of ten years of earnings per share, each restated in present-day dollars using annual CPI-U before averaging. Averaging nominal earnings across a decade would understate the earnings base and make every stock look cheaper than it is.
The percentile is cross-sectional— a rank within the scanned universe, not against a historical distribution. A stock in the cheapest 30% is cheap relative to today's index, which in an expensive market is not the same as cheap outright.
Reported EPS is not always usable as published. The data provider back-adjusts stock splits inconsistently, so every year is restated onto the share basis the current price is quoted on. Where that restatement cannot be made coherently, the stock gets no CAPE at all rather than a wrong one — which is why roughly a fifth of the index can never produce a signal.
F-Score — quality
Piotroski's nine binary tests across profitability, leverage and efficiency. Each is evaluated independently and may come back undetermined when a filing omits the input. An undetermined test is recorded as such and never counted as a pass: scoring a missing figure as a pass inflates the score, and treating an absent debt line as “no debt” would award the leverage point outright. Any incomplete score holds the quality leg at neutral.
Ratios use end-of-period total assets where the original paper uses a two-year average, so scores may differ by a point from published figures.
Point & Figure — timing
Six-percent logarithmic boxes on a three-box reversal, built from weekly highs and lows rather than closes — an intra-week penetration is exactly what the chart exists to capture. Log scaling keeps the box a constant percentage at every price level. A breakout counts only while it belongs to the column currently being drawn; one from nine columns ago is history, not timing.
The box was 2% until testing showed why it should not be. A 2% box on a three-box reversal triggers on roughly a 6% move, which plays out over the few weeks across which stock returns tend to reverse— the same window momentum research deliberately excludes. Measured point-in-time, the leg's cost from firing at the wrong moments fell at every one of nine box widths tested, from -1.06% per quarter at 2% to -0.57% at 6% and about -0.46% beyond 10%.
Six percent rather than the ten or twelve the timing curve alone would favour, because coverage binds first. A wider box needs more price history to draw two columns, and the data feed supplies two years. At 12% roughly a sixth of the index could never produce a signal at all — on top of the fifth that already has no computable CAPE. Six percent removes just under half the timing cost while leaving 98% of the index scoreable.
What a backtest found
The rule was tested point-in-time over 2010–2025 across 492 companies and 81,899 ticker-observations. Statements were used only from the date they were actually published, entry was priced at the following week's open rather than the decision bar, and earnings were deflated to the CPI base of the simulated year.
| Cohort | n | Mean 13-week P&L |
|---|---|---|
| All three agree (Buy) | 961 | +1.90% |
| All three agree (Sell, as a short) | 118 | -5.96% |
| Every scanned stock | 81,899 | +3.80% |
The Sell row is the profit of the short position, not the move in the stock: these names rose about 6%, which costs a short about 6%. Reporting both cohorts as stock moves — as an earlier version of this page did — makes a losing short read as an outperforming pick.
The agreement rule underperformed by about two points a quarter, and the sell signal did not work in the other direction either. The result held across every hold period from one week to one year.
Where the shortfall comes from
A gap against the whole universe conflates two different failures: picking worse companies, and picking good ones at worse moments. Splitting them requires comparing the flagged cohort only against stocks scored on the same dates. What remains is selection; the rest is timing.
| Hold | Selection | Timing | 95% interval on selection |
|---|---|---|---|
| 4w | -0.15% | -0.47% | [-0.63%, +0.35%] |
| 8w | -0.19% | -0.92% | [-1.00%, +0.62%] |
| 13w | -0.65% | -1.25% | [-1.72%, +0.41%] |
| 26w | -1.72% | -0.76% | [-3.02%, -0.44%] |
| 52w | -3.25% | -0.90% | [-5.13%, -1.43%] |
Out to about two months the damage is almost entirely timing, and the selection term cannot be distinguished from zero — its interval spans both signs. That is the Point & Figure leg doing what it is built to do: a breakout requires a price that has already risen, so the rule systematically buys after a run. Only at six months and beyond does a genuine stock-selection shortfall emerge that survives the interval.
Intervals resample whole months, not individual positions. The 81,899 rows are not 81,899 independent facts — within any month every holding moves with the market, so they are closer to 192. Treating them as independent would shrink every interval by roughly the square root of the cohort size and turn noise into a finding.
2012–2025 strongly favoured expensive, fast-growing companies — precisely what a cheap-valuation screen excludes by construction — so this is one market regime rather than a verdict on the methods themselves. It is also a single historical path, on large-cap US equities, with no transaction costs and no dividends.
Two biases work in opposite directions and neither is corrected. The universe is today's index constituents, so companies dropped over the window are missing and several present were added because they did well — that inflates the benchmark more than the flagged cohort, meaning the true gap is probably smaller than measured. Against that, only the latest version of each filing is stored, so a restated figure a contemporary investor could not have seen may be in use.
We publish this because it is what the evidence shows. The screen is offered as a transparent, checkable way to find where three published methods coincide — not as a strategy demonstrated to beat the market.
Freshness and coverage
The universe is rescanned on weekdays after the US close. Every run records how many companies it analysed and how many it could not, and both numbers are shown on the screen. A run that fails leaves the previous one in place rather than publishing a partial scan, and the screen always states when it was computed.