Discovery Engine

A private research engine that runs locally: it screens US stocks on point-in-time fundamentals and price behaviour, then simulates what its rules would have done. This page reports what data it holds and how the tests came out — nothing here is a recommendation.

Snapshot generated 2026-09-05 14:12 UTC

What data we have

Dataset	Rows	Tickers	Covers
Fundamentals (SEC XBRL)7,644 companies · 48 metrics	18,513,480	7,295	2009-04-15 → 2026-07-27
Daily prices — deep history	19,470,643	6,392	1962-01-02 → 2026-08-28
Daily prices — recent	2,916,430	6,440	2024-07-26 → 2026-09-04
Prices with raw + adjusted + dividends	1,439,965	394	1980-03-17 → 2026-09-04
Stock splits	25,674	17,515	2005-01-03 → 2026-12-17
Cash dividends	56,458	26,346	2026-04-15 → 2030-12-13
Forward analyst estimates	98,785	3,815	2026-08-10 → 2027-07-20

How far back we can test

Prices reach back decades, but company filings only became machine-readable from 2009. Rules that read financials can only be tested where enough companies had filed — which is what the right-hand column shows.

Year	Tickers priced	With financials
2000	296	none
2005	1,936	none
2010	2,298	463
2013	2,647	3,477
2016	3,082	4,058
2020	3,890	5,084
2024	5,318	6,461

Latest screen · 2026-07-28

9,297

Universe

2,407

Survived

6,890

Eliminated

illiquid: 2,329ps_extreme: 233

The tests

Each block is one experiment. The line under every heading is its result, so you can read the outcomes without opening anything; expand one for the chart, the method and the caveats. These are tests, not a track record.

Two ways to avoid overpaying2021–2026Same money, same schedule, same ranking — differing in one decision: how each judges a stock too expensive. Both lost (+29.6% and +40.7%) while SPY made +77.0%.

Both models start on the same day with the same money, buy and sell on the same monthly schedule, hold the same number of companies, and rank them the same way. They differ in exactly one decision — how they decide a stock is too expensive — so the gap between the two lines is what that one decision was worth.

Cumulative return. Both models track the market closely, so the difference is easier to read in the other view.

Model A · relative

Judges 'expensive' against whatever is on offer that month and throws out the dearest slice of it. Always finds something to buy, because the bar moves with the era it is in.

Return: +29.6%
Per year: +5.3%
Worst fall: -40.8%
Return per unit of risk: 0.18
Round trips: 671
Went up: 53%

Model B · absolute

Judges 'expensive' against fixed bounds that do not move with the crowd, and additionally wants a real margin, no reverse split and little dilution. Holds fewer names — or cash — when nothing clears the bar. The exact bounds in force are listed beside it.

Return: +40.7%
Per year: +7.1%
Worst fall: -39.3%
Return per unit of risk: 0.25
Round trips: 668
Went up: 54%

Rebased to 0% on 2021-07-29. Every trade is decided on one day and filled at the next session's opening price, with 25 basis points of cost charged each way. The chart is sampled weekly; the figures above come from the daily series. Both models were tested only on companies that still exist today, which flatters them — see Known gaps.

Dividends are only partly counted. The price history behind every line carries a median of about a third of the dividends actually paid, measured against a known-good source. It applies to the models and the index alike, so the comparison is fair, but each line understates its true total return — SPY's real figure over this window is nearer +79.5% than the +77.0% drawn here. Because value shares yield more than the index, the shortfall is larger for the two models than for the market, so this works against them, not for them.

Model A · relativenow +29.6%

Current rules: at least $50M traded a day · pe_ratio > 0, ps_ratio > 0

Earlier versions of these rules

Kept so the change is attributable, not just announced. The settings column is what explains the difference.

Run	Return	Worst fall	Settings that differed
2026-08-01 v1	+29.6%	-40.8%	≥ $50M/day traded · pe_ratio > 0, ps_ratio > 0
2026-08-01 v1	+29.6%	-40.8%	≥ $50M/day traded · pe_ratio > 0, ps_ratio > 0
2026-08-01 v1	+29.6%	-40.8%	pe_ratio > 0, ps_ratio > 0· liquidity floor not recorded for this run
2026-08-01 v1	-23.5%	-49.2%	pe_ratio > 0, ps_ratio > 0· liquidity floor not recorded for this run
2026-07-31 v1	-21.3%	-49.2%	pe_ratio > 0, ps_ratio > 0· liquidity floor not recorded for this run

A few winners and a few losers

8 positions picked at random — an equal number of gains and losses, not the biggest of either, and drawn from both sold and still-held lots. Prices are adjusted for splits and dividends.

Stock	Bought	At	Sold	At	Held	Return
MU	2022-11-29	$55.71	2022-12-29	$49.40	30d	-11.5%
HUN	2023-06-01	$23.78	2023-07-03	$27.04	32d	+13.4%
M	2023-08-02	$16.10	2023-08-31	$12.34	29d	-23.5%
LW	2024-10-02	$63.00	2024-10-31	$77.07	29d	+22.0%
FIVE	2025-03-06	$83.87	2025-04-04	$53.63	29d	-36.2%
RHI	2026-01-06	$26.57	2026-02-05	$32.94	30d	+23.7%
FOUR	2026-01-06	$63.84	2026-02-05	$58.45	30d	-8.7%
LRN	2026-03-09	$85.46	2026-07-28	$95.88	141d	+11.9%

Measured against SPY. Trading costs are charged on every one of these.

Model B · absolutenow +40.7%

Current rules: at least $50M traded a day · dilution_1y <= 0.1, pe_ratio <= 40, profit_margin >= 0.05, ps_ratio <= 10, reverse_splits <= 0

Earlier versions of these rules

Kept so the change is attributable, not just announced. The settings column is what explains the difference.

Run	Return	Worst fall	Settings that differed
2026-08-01 v2	+40.7%	-39.3%	≥ $50M/day traded · dilution_1y <= 0.1, pe_ratio <= 40, profit_margin >= 0.05, ps_ratio <= 10, reverse_splits <= 0
2026-08-01 v2	+40.7%	-39.3%	≥ $50M/day traded · dilution_1y <= 0.1, pe_ratio <= 40, profit_margin >= 0.05, ps_ratio <= 10, reverse_splits <= 0
2026-08-01 v2	+40.7%	-39.3%	dilution_1y <= 0.1, pe_ratio <= 40, profit_margin >= 0.05, ps_ratio <= 10, reverse_splits <= 0· liquidity floor not recorded for this run
2026-08-01 v1	-11.8%	-42.7%	dilution_1y <= 0.1, pe_ratio <= 20, profit_margin >= 0.05, ps_ratio <= 3, reverse_splits <= 0· liquidity floor not recorded for this run
2026-07-31 v1	-11.6%	-42.6%	dilution_1y <= 0.1, pe_ratio <= 20, profit_margin >= 0.05, ps_ratio <= 3, reverse_splits <= 0· liquidity floor not recorded for this run

A few winners and a few losers

8 positions picked at random — an equal number of gains and losses, not the biggest of either, and drawn from both sold and still-held lots. Prices are adjusted for splits and dividends.

Stock	Bought	At	Sold	At	Held	Return
CHD	2021-08-30	$81.51	2021-09-29	$81.94	30d	+0.3%
BBWI	2022-12-29	$41.35	2023-03-02	$39.45	63d	-4.8%
SWN	2023-01-31	$5.40	2023-03-02	$5.32	30d	-1.7%
ALB	2023-10-02	$169.19	2023-10-31	$126.67	29d	-25.3%
GIS	2024-06-03	$68.69	2024-07-03	$63.01	30d	-8.5%
SJM	2024-10-31	$113.99	2024-12-02	$118.00	32d	+3.3%
CRM	2025-12-04	$243.69	2026-01-06	$255.20	33d	+4.5%
OWL	2026-07-09	$9.27	2026-07-28	$9.81	19d	+5.6%

Measured against SPY. Trading costs are charged on every one of these.

Buy every 10% drop, sell at a doublecontrolEvent-driven, no ranking, no valuation test — shares no machinery with the models above. +10.1% on the money put in, against +45.9% for the same cash into SPY. Two opposite rules, the same verdict.

A control, built to test the engine rather than to make money: buy a fixed amount of any liquid stock that falls 10% in a day, buy again every time it falls again, and sell a lot only once it has doubled. Nothing is ever cut.

The money arrives as trades happen, so this is a return on everything paid in, not the growth of one opening stake — a different measure from the first chart, which is why it is not on it. The fair comparison is the same money going into the index on the very same days, and that is the dashed line.

The rule+10.1%Same money into SPY+45.9%

Put in

$49.0M

49,042 buys of $1,000

Ended worth

$54.0M

+10.1% on the money in

Same money in SPY

+45.9%

the like-for-like comparison

Why the win rate is a trap here. A position can only be closed by doubling, so the 16,828 sales were almost all winners by construction — the figure measures the rule, not the skill. Everything that went wrong is still being held: 32,214 open positions that cost $32.2M are worth $16.8M (-48.0%). Selling winners and holding losers is the whole result.

A few winners and a few losers

8 positions picked at random — an equal number of gains and losses, not the biggest of either. Both sold and still-held are drawn from deliberately. A lot can only be SOLD here by doubling, so every sold row is a gain and all the losses are still held. Sampling closed trades alone would therefore have shown one side only. Prices are adjusted for splits and dividends, so they are the comparable series rather than the number printed on the ticker that day.

Stock	Bought	At	Sold	At / now	Held	Return
XOMA	2022-06-13	$17.30	2025-09-03	$34.53	1178d	+98.6%
NXDR	2022-11-10	$2.19	still held	$2.30	1356d	+4.8%unrl.
EJH	2024-05-21	$8,000.00	2024-07-08	$18,750.00	48d	+133.2%
FFAI	2024-06-21	$2,480.40	still held	$4.63	767d	-99.8%unrl.
CMTG	2025-04-11	$2.46	still held	$2.12	473d	-14.0%unrl.
NNNN	2025-11-03	$16.82	2025-12-04	$37.40	31d	+121.2%
AMKR	2026-06-08	$68.40	still held	$45.69	50d	-33.4%unrl.
PMI	2026-06-08	$0.26	still held	$0.09	50d	-66.0%unrl.

Earlier versions of these rules

Kept so the change is attributable, not just announced. The settings column is what explains the difference.

Run	Return	Worst fall	Settings that differed
2026-08-01 v1	+10.1%	-35.5%	≥ $50M/day traded
2026-08-01 v1	+9.8%	-35.5%	liquidity floor not recorded for this run
2026-07-31 v1	+9.8%	-35.5%	liquidity floor not recorded for this run
2026-07-31 v1	+9.8%	-35.5%	liquidity floor not recorded for this run
2026-07-31 v1	+9.8%	-35.6%	liquidity floor not recorded for this run

Buy every 10% jump, cut at -10%mirror controlThe same machinery with every sign flipped: buys strength, cuts losers, rides winners. +8.3% on the money put in, against +44.6% for the same cash into SPY. Two opposite rules, the same verdict.

The exact opposite rule, run through the same machinery: buy any liquid stock that RISES 10% in a day, never take a profit, and sell only once a lot sits 10% below what it cost. It cuts its losers and rides its winners — the textbook discipline, and the reverse of what the model above does.

The rule+8.3%Same money into SPY+44.6%

Put in

$66.5M

66,479 buys of $1,000

Ended worth

$72.0M

+8.3% on the money in

Same money in SPY

+44.6%

the like-for-like comparison

Why the win rate is a trap here. A position can only be closed by falling 10%, so the 59,810 sales were losers by construction — the figure measures the rule, not the skill. Everything that went right is still being held: 6,669 open positions that cost $6.7M are worth $18.4M (175.9%). Cutting losers and riding winners is the whole result — the exact reverse of the model above, and it lands in almost the same place.

One number here is assumed, not measured. When a stop fires, this model books the sale at exactly −10%. In reality a stock often closes far below that, and no order could have filled at a price it never traded at. Running the identical rules while selling at the price that actually traded turns +8.3% into +4.5% — so roughly half of what this model appears to earn comes from that one assumption.

A few winners and a few losers

8 positions picked at random — an equal number of gains and losses, not the biggest of either. Both sold and still-held are drawn from deliberately. A lot can only be SOLD here by hitting its stop, so every sold row is a loss and all the gains are still held. Sampling closed trades alone would therefore have shown one side only. Prices are adjusted for splits and dividends, so they are the comparable series rather than the number printed on the ticker that day.

Stock	Bought	At	Sold	At / now	Held	Return
WRAP	2021-08-06	$8.14	2021-08-19	$7.33	13d	-10.4%
NU	2022-06-22	$3.77	still held	$14.68	1497d	+288.4%unrl.
SBEV	2023-10-06	$111.20	2023-10-25	$100.08	19d	-10.4%
HMR	2025-03-19	$3.44	2025-03-24	$3.10	5d	-10.4%
CMI	2025-04-10	$288.04	still held	$641.60	474d	+122.2%unrl.
SVRA	2025-10-21	$4.09	still held	$5.33	280d	+30.0%unrl.
KZIA	2026-04-23	$11.46	still held	$12.69	96d	+10.5%unrl.
VECO	2026-06-10	$67.45	2026-07-02	$60.71	22d	-10.4%

Earlier versions of these rules

Kept so the change is attributable, not just announced. The settings column is what explains the difference.

Run	Return	Worst fall	Settings that differed
2026-08-01 v1	+8.3%	-11.8%	≥ $50M/day traded
2026-08-01 v1	+8.3%	-11.8%	liquidity floor not recorded for this run
2026-07-31 v1	+8.3%	-11.8%	liquidity floor not recorded for this run

Every run, side by side12 sets of rulesOne row per set of rules, showing its most recent run and how many times it has been simulated. 1 of 12 beat the index they were measured against.

Each row is the same rules run over a past period, buying and selling on the dates they would have fired. Trading costs are charged. Returns include only part of the dividends paid — see the note below the chart. One row per set of rules, showing its most recent run; “Runs” counts how many times that set was simulated, because a result picked from many attempts is weaker evidence than one from a single attempt.

Strategy	Period	Return	vs index	Worst fall	Trades	Runs
pop_stop_at_close	2020-07-29 → 2026-07-28	+4.5%	+44.6%SPY	-13.7%	59,810	5
pop_stop_or_hold	2020-07-29 → 2026-07-28	+8.3%	+44.6%SPY	-11.8%	59,810	4
dip_double_or_holdon cash paid in	2020-07-29 → 2026-07-28	+10.1%	+45.9%SPY	-35.4%	16,828	12
deep_neglect_price_only	2013-01-02 → 2026-07-28	+345.5%	+283.3%IWM	-46.2%	280	9
deep_value_neglect	2013-01-02 → 2026-07-28	+216.4%	+283.3%IWM	-63.1%	280	9
value_momentum	2021-07-29 → 2026-07-28	+50.9%	+77.0%SPY	-24.2%	380	2
quality_neglect	2021-07-29 → 2026-07-28	+27.2%	+77.0%SPY	-46.4%	620	2
absolute_cheap_value_ranked	2021-07-29 → 2026-07-28	+54.8%	+77.0%SPY	-31.3%	246	2
absolute_cheap_fixed	2021-07-29 → 2026-07-28	+33.3%	+77.0%SPY	-33.2%	474	2
relative_cheapest_quartile	2021-07-29 → 2026-07-28	+34.0%	+77.0%SPY	-40.9%	399	2
absolute_value_monthly	2021-07-29 → 2026-07-28	+40.7%	+77.0%SPY	-39.3%	668	15
relative_value_monthly	2021-07-29 → 2026-07-28	+29.6%	+77.0%SPY	-40.8%	671	15

Paying for the searchoverfitting checkTry enough rules and the best one looks skilled by luck alone. With 12 tested, that luck is worth +0.46. Nothing clears the 0.95 bar — the best reaches 0.57.

Try enough sets of rules against the same history and the best one looks skilled purely by luck. That bar is measurable: with 12 sets tested, the best of that many worthless strategies would still have posted a return-to-risk score of +0.46. Anything at or below that line is explained by the search alone. The last column is the probability the result is real after paying that cost — above 0.95 is the usual bar.

Strategy	Return per unit of risk	Probability it is real
absolute_cheap_fixed	+0.35	0.397
absolute_cheap_value_ranked	+0.48	0.512
absolute_value_monthly	+0.38	0.430
deep_neglect_price_only	+0.51	0.570
deep_value_neglect	+0.42	0.439
dip_double_or_hold	-0.21	0.067
pop_stop_at_close	-0.35	0.035
pop_stop_or_hold	+0.12	0.225
quality_neglect	+0.31	0.368
relative_cheapest_quartile	+0.35	0.399
relative_value_monthly	+0.32	0.378
value_momentum	+0.53	0.555

The count of tests is a floor, not a total — it counts only the runs that were recorded, never the thresholds nudged and abandoned along the way. So the real bar is higher than the one shown.

Known gaps

Listed because a result is only worth as much as the honesty about what is missing from it.

Survivorship — the companies that failedthe main limitation

Every simulation here was run on companies that still exist. The price archive carries almost nothing that went to zero, so the worst outcomes are missing and every result is flattered. Being fixed slowly: 396 delisted companies' price histories collected so far, capped at ~450 a month by the free data tier.

Linking dead companies to their financialsmostly solved

When a company delists, public sources drop the link between its ticker and its regulatory filings. Archived copies of the SEC's own ticker file recovered 7,264 of those links; 349 companies remain unmatched, mostly ones that delisted before 2017 when the archives begin.

Dividends are only partly countedmeasured, not fixed

The price history behind every result carries a median of about a third of the dividends actually paid. It applies to the strategies and the index alike, so comparisons stay fair, but each figure understates its true total return — and by more for high-yield value shares than for the market.

Costs are simplifiedknown

Every trade is charged a flat 25 basis points each way, regardless of how small or illiquid the company is, and there is no model of market impact or tax. Real trading would cost more, especially in the smallest names.

Sector codesdone

Needed because bank profit margins are not measurable from XBRL, and the P/E rule reads margin.

Measures it computes (62)referenceEvery quantity the screen can read about a company, with the version of the definition that produced it.

accrualsv1 · ratio

(TTM net income - TTM operating CF) / assets; high = suspicious

assetsv1 · USD

Total assets, latest reported

assets_currentv1 · USD

Current assets

buyback_yieldv1 · ratio

TTM buybacks / market cap

cashv1 · USD

Cash and equivalents, latest reported

cash_conversionv1 · ratio

TTM operating cash flow / TTM net income

cash_to_assetsv1 · ratio

Cash / total assets

current_ratiov1 · ratio

Current assets / current liabilities

debt_to_equityv1 · ratio

Long-term debt / equity

decline_daysv3 · days

Days since LAST at that peak

dilution_1yv3 · ratio

Share count growth over ~1 year, split-adjusted

dividend_yieldv1 · ratio

TTM dividends paid / market cap

dollar_volumev2 · USD

Median daily traded value (tradability)

drawdownv2 · ratio

Fall from trailing-252d peak

earnings_growth_1yv1 · ratio

TTM net income growth; null off a negative base

equityv1 · USD

Stockholders equity, latest reported

fcf_marginv1 · ratio

TTM free cash flow / TTM revenue

fcf_yieldv1 · ratio

TTM free cash flow / market cap

forward_epsv1 · USD

Consensus EPS, next 4 quarters, as visible on as_of

forward_pev1 · ratio

Price / forward 4-quarter consensus EPS

goodwillv1 · USD

Goodwill on the balance sheet

goodwill_to_assetsv1 · ratio

Goodwill / total assets (acquisition risk)

gross_marginv2 · ratio

TTM gross profit / TTM revenue

interest_coveragev1 · ratio

TTM operating income / TTM interest expense

inventoryv1 · USD

Inventory, net

inventory_to_revenuev1 · ratio

Inventory / TTM revenue

liabilitiesv1 · USD

Total liabilities, latest reported

liabilities_currentv1 · USD

Current liabilities

long_term_debtv1 · USD

Long-term debt (total, else noncurrent)

market_capv3 · USD

RAW price x as-reported shares

net_debtv1 · USD

Long-term debt minus cash

ocf_growth_1yv1 · ratio

TTM operating cash flow growth

operating_marginv1 · ratio

TTM operating income / TTM revenue

pe_ratiov3 · ratio

Market cap / TTM net income (positive only)

ppe_netv1 · USD

Property, plant and equipment, net

price_to_bookv1 · ratio

Market cap / stockholders equity

profit_marginv2 · ratio

TTM net income / TTM revenue

ps_ratiov3 · ratio

Market cap / TTM revenue

public_floatv1 · USD

Public float as stated on the cover page

rd_intensityv1 · ratio

R&D / TTM revenue

receivablesv1 · USD

Accounts receivable, net current

receivables_to_revenuev1 · ratio

Receivables / TTM revenue

retained_earningsv1 · USD

Retained earnings / accumulated deficit

revenue_growth_1yv1 · ratio

TTM revenue vs the same measure a year earlier

reverse_splitsv1 · count

Reverse splits in trailing year (distress)

roav1 · ratio

TTM net income / total assets

roev1 · ratio

TTM net income / equity

sbc_to_revenuev1 · ratio

Share-based comp / TTM revenue (real dilution cost)

shares_outv3 · count

Shares outstanding, as reported

ttm_buybacksv1 · USD

Trailing 12m share repurchases (absolute)

ttm_capexv1 · USD

Trailing 12m capital expenditure (absolute)

ttm_dividends_paidv1 · USD

Trailing 12m dividends paid (absolute)

ttm_free_cash_flowv1 · USD

TTM operating cash flow - capex, period-aligned

ttm_gross_profitv2 · USD

Trailing 12m gross profit

ttm_interest_expensev1 · USD

Trailing 12m interest expense (absolute)

ttm_net_incomev2 · USD

Trailing 12m net income

ttm_operating_cash_flowv1 · USD

Trailing 12m cash from operations

ttm_operating_incomev1 · USD

Trailing 12m operating income

ttm_rdv1 · USD

Trailing 12m research and development

ttm_revenuev2 · USD

Trailing 12m revenue, point-in-time

ttm_sbcv1 · USD

Trailing 12m share-based compensation

volume_trendv2 · ratio

30d mean volume / prior mean volume

Research output only. Nothing on this page is investment advice, and no result here is a prediction of future returns.