THE FACTOR LAB HANDBOOK

Every day, a computer reads the financial reports and price history of roughly 6,600 US stocks and ranks them on one leaderboard. Top of the leaderboard = the most evidence in the stock's favor. That's it. This handbook walks through exactly how that evidence is assembled, what every number and chip on screen means, and what the system is deliberately not doing.

If a word has a dotted underline, it is a Ticker-level technical term — click it to see what it means without leaving the page. A complete, searchable index of every term lives in the glossary section.

Behind the page is an ordered chain of jobs, re-run automatically every day via GitHub Actions. An orchestrator enforces the order and checks data integrity so stale or partial data can never silently corrupt your rankings.

Everything is Point-in-time: each logged signal uses only information that existed at that moment. That discipline is what makes the track record trustworthy.

The Rankings tab is a table of the whole universe, ranked by one number: the Composite score. Here is what every column means.

• Rank — today's position on the leaderboard of ~6,600 scored US stocks. #1 has the strongest overall evidence right now.
• Composite — the one number everything is ranked by (0–100). It blends the five factor scores, each measured against the stock's own Sector-neutral comparison group, then applies safety Safety haircuts. Higher = more evidence in the stock's favor.
• Factor mix — what is driving the score. A contribution bar shows the share of each factor: green = value, blue = quality, amber = momentum, violet = low volatility, pink = revisions. Longer segment = bigger contribution.
• Band — what the rank means in practice (Research Now, Watchlist, Monitor, Pass). A red chip means the stock is Vetoed outright — the reason is written on the chip.
• Market cap — Market cap, the price of the whole company (share price × shares).
• RS2 rank / stance / conviction / action — the independent AI read. See the RS2 section.
• Δ pctl — how much the quant engine and the AI disagree, in percentile points. Big gaps are the interesting rows: one of them is wrong.
• DCF gap — the Expectations gap (DCF gap): the growth the price requires vs the growth the company has actually delivered. See the DCF section.

Each stock is graded on five traits that have historically predicted returns. Crucially, every grade is Sector-neutral — a supermarket competes with supermarkets, not with software companies. Otherwise “high momentum” would just mean “is a tech stock.”

Why every ingredient counts equally

Think of judging a Factor decathlon: you could try to guess which event matters most, but decades of research show those guesses backfire — the “perfect” weights found in past data almost never work on future data. So each of the five ingredients counts exactly the same (Equal weighting (1/N)). Boring, humble, and it works better. The site still Rank IC-measures each factor's predictive power every month as a diagnostic — it just never lets a short sample steer the engine.

The composite percentile is cut into four practical buckets, or Bands:

• RESEARCH NOW — top 3%. Worth your research time today.
• WATCHLIST — top 10%.
• MONITOR — top 30%.
• PASS — the rest.

Vetoes — hard disqualifiers

Some stocks are disqualified no matter how good the score looks — think of a house with beautiful photos that failed the structural inspection. A Veto is applied before scoring, so a vetoed name gets no composite at all. Reasons include: the reverse engine's safety checks (reverse_engine_reject), both forensic alarms firing together (Beneish M-score + high Accruals), Share dilution / issuance from heavy share issuance, or (in the AI lens) a hard avoid/sell verdict. The reason is written on the red chip.

Safety haircuts

Between the raw score and the final rank, three multiplicative Safety haircuts are applied: survivability = 0.7 + 0.3 × (survivability/100); data quality = min(1, 0.8 + 0.04 × dq); and forensic = 0.85 if a single Beneish or accruals alarm fired (both firing is a veto, not a haircut). The result is re-ranked, so fragile or suspicious names drop without being thrown out.

Every stock price silently makes a promise about future growth. The Reverse DCF extracts that promise as a number — the Implied growth the market is charging you for — by solving (via Bisection) for the growth rate that makes a standard Discounted cash flow (DCF) equal the current price.

The DCF gap column compares that promise with reality: Implied growth minus Demonstrated growth (the last 5 years of revenue/FCF growth from SEC filings), in percentage points.

• Green / negative — the price promises LESS than the company has proven. A potential bargain: you are being paid not to believe the growth story.
• Amber / positive — the price needs an acceleration nobody has demonstrated yet. You have to believe a story.

At the top of the Rankings tab is one switch that decides whose ranking you see:

• Quant — the deterministic factor engine. Pure math over financial statements and prices. No AI involved. This is the original, unchanged view.
• RS2 LLM — the local AI analyst's own list, built by reading each company's actual filings and writing an independent verdict.
• Compare — both side by side, biggest disagreements first. A disagreement percentile (Δ pctl) is computed per name; big gaps are where one engine is wrong.

The AI applies after the quant bands are set: it can promote, demote, or veto names, producing a parallel ranking. The quant baseline is never overwritten — the two lists are both shown so you can see where they disagree.

RS2 is a LLM (large language model) that reads each company's actual SEC filings and writes an independent verdict — like getting a second doctor's opinion. For every name it produces:

• Stance — undervalued / fair / overvalued (a colored pill).
• Conviction — how confident it is, 0–15.
• Action — buy / hold / reduce / avoid… an opinion for research, never an order.
• Its own DCF read, whose Intrinsic value is anchored to the analyst consensus band and de-forwarded to Present value / discounting — so the Margin of safety measures cheapness today, not a 12-month price target.

The AI Research Now gate

The AI's Research Now list keys off RS2's structured signals — margin of safety and entry timing — in two tiers: deep value (MoS ≥ 30%) earns Research Now at any conviction; moderate value (MoS ≥ 15%, or a genuine fresh buy) additionally needs conviction ≥ 9.5. Bearish calls are demoted out of Research Now; a hard avoid/sell is vetoed.

When a name leaves the quant Research Now list, RS2 writes an Exit review — a hold/trim/sell call for current holders, shown as an amber “LLM EXIT” chip.

Instead of showing a flattering Backtest, the system Paper tradings its own picks every single day with real prices and real Transaction costs (bps), and the record is append-only — it can never be edited. If the machine is wrong, this page will say so, publicly and permanently. That's the point.

The portfolios

• plan — the value core: Kelly criterion-sized, ~50% cash.
• plan2 — the hybrid: value core + Quality sleeve, ~78% invested, holds the expensive leaders.
• equal — equal-weighting every Research Now name (pure stock-picking test).
• mine — your saved My Portfolio holdings, Unitization-measured like a fund.

How to read it

• plan vs plan2 — if plan2 wins, paying up for quality leaders beat the value discipline this period; if plan wins, the discipline (and cash) paid off.
• plan vs equal — the sizing machinery adds value if plan beats equal weighting.
• equal vs IWM — the stock selection itself works if the picks beat the small-cap Benchmark.
• mine vs plan — your own deviations cost money: that's the Behavior gap.
• “Sold too early” flags — exits that kept rising. A recurring pattern there means the exit rule needs work.

Sharpe ratio and CAGR appear only after enough days of live data; early on this page is deliberately boring. A what-if overlay lets you re-cost every trade at your own commission rate to see the drag of fees.

The Portfolio tab has two very different halves. Read the labels carefully:

My Portfolio (top) — your actual holdings

Enter your ACTUAL holdings (saved only in this browser). Each is checked against the model: a quarter-Kelly criterion suggested size, an over/under-weight verdict, and loud flags if a holding is Vetoed or outside coverage. The “mine” ledger on Track Record uses these, unitized like a fund.

Suggested plan (below) — NOT your portfolio

A machine-built allocation from the Research Now list, with a Value core / Hybrid toggle:

• Value core (plan) — quarter-Kelly sizing: Edge = the expectations gap closing over ~3 years; risk = Volatility (σ); f = 0.25 × edge/risk², capped 5%. Forensic flags flags halve size; GPR 2–3 and insider selling shrink it; sector 25% / theme 30% caps; the rest stays Cash (in the plan) (often ~50%).
• Hybrid (plan2) — the same value core PLUS a Quality sleeve that buys the top-ranked names REGARDLESS of valuation gap (capped ~35% of Book value), so it holds the expensive leaders the core refuses and deploys the idle cash (~78% invested). Sleeve rows are tinted pink.

Macro de-risk

Macro flags are warning lights from FRED data. If 2+ fire, every suggested size halves automatically (Macro de-risk) — shown as a loud amber banner on the tab.

Overlays are context chips, never additive score. They exist to shrink positions, demand bigger margins of safety, or question your thesis:

• GPR 0–3 — GPR (geopolitical exposure) tagged from the company's actual business profile (revenue geography, supply chains, regulation, sanctions). Never a buy/sell signal; at level 3 it shrinks position sizes and demands a bigger margin of safety.
• ▲/▼ INSIDERS — Informed demand: insiders net-buying while short sellers retreat (▲, confirming) or insiders selling into elevated Short interest (▼, interrogate the thesis). Confirmation or warning only.

Forensic flags

The forensic battery — Beneish M-score M-score, Sloan Accruals, net issuance, and the fundamentals battery — produces the Forensic flags flags shown on the plan rows. A single alarm is a 0.85 haircut; the pair firing together is a Veto.

A Theme is a market narrative a stock belongs to — AI, semiconductors, biotech, and so on. Theme membership and scores ride along for orientation and for crowding warnings (late-cycle theme crowding is a risk signal), but they never add to the composite.

Scoring pipeline — exact mechanics

Universe: every name scored by the reverse engine (~6,600 US listings). Per sub-metric:Winsorize at the 1st/99th percentile within sector, then Z-score within sector. Factor z = mean of that factor's available sub-metrics. Composite z = weight-renormalized sum over available factors (missing factors drop out and remaining weights rescale; value, quality, and momentum are required — a name missing any of them is marked insufficient_factors rather than scored on partial data).

Composite z → cross-sectional Percentile (0–100) → three multiplicative Safety haircuts (survivability, data quality, forensic) → re-ranked → final percentile sets the Band: ≥97 research_now, ≥90 watchlist, ≥70 monitor, else pass.

Factor construction — sub-metrics and sources

Value = mean z of four yields, all computed from the latest fiscal year of SEC-filed fundamentals against current market cap:Free cash flow (FCF) yield (FCF/mcap), Owner earnings yield ((NI + D&A − capex)/mcap),EBIT yield (operating income/Enterprise value (EV)), and Earnings yield (NI/mcap — broadest coverage, rescues filers with missing capex/D&A/op-income tags).

Quality = mean z of: Revenue quality (reverse-engine score), Gross margin stability (−stdev across ≥4 fiscal years), negative Accruals (−accruals ratio), and Piotroski F-score (both from the forensic battery).

Momentum = mean z of the 12-1 skip-month return and 52-week-high proximity. Monthly closes.

Low volatility = z of −σ(monthly returns), minimum 12 observations; the Annualized volatility is exported per name and feeds Kelly criterion sizing.

Revisions = mean of two 0–1 parts: normalized EPS trajectory slope (clamp(slope, −1, 1)+1)/2, and analyst structured score/100 — scaled to 0–100 then re-centred to a z-like scale via (score−50)/25.

Weights

Equal 0.20 × 5 (scheme equal_weight_robust5). Rank IC per factor is measured monthly but writes a drift diagnostic only — measured IC never steers the weights (DeMiguel, Garlappi & Uppal 2009: estimated weights rarely beat 1/N out of sample).

Reverse DCF — exact method

The valuation models solve by Bisection for the growth rate that makes a standard DCF equal the CURRENT price — the growth the market is charging you for. The Expectations gap (DCF gap) (shown as “DCF gap”) = implied growth − demonstrated growth, where demonstrated = the last 5 years of revenue/FCF growth from SEC filings, in percentage points. The reverse engine layers archetype classification (A–F) and survivability/data-quality scoring on top, producing the safety inputs the Factor Lab consumes.

Veto rules (exact)

reverse_engine_reject = reverse-engine band ∈ {Excluded, Reject, Reject-tier}· forensic_pair = Beneish M-score elevated AND accruals high (single alarm = 0.85 haircut instead) · heavy_issuance = HEAVY_ISSUANCE flag, waived for archetypes E/F where issuance is the expected financing mode. In the AI lens, a hard avoid/sell verdict is also a veto (llm_reject).

Missing data is null, never silently safe

The old placeholder Z/M-scores are gone. If a metric is missing, it is null — and the scoring pipeline either marks the name insufficient or applies the data-quality haircut. A gap is never quietly treated as a pass.

Two independent honesty loops keep the machine honest:

• Forward-logged signals — every factor signal is logged at the moment it is made (Point-in-time, append-only) and later measured against what actually happened, including delisted names (no Survivorship bias).
• Paper-traded portfolios — the Track Record page trades plan / plan2 / equal / mine daily with real prices and costs, benchmarked against IWM and SPY. Returns and Alpha / excess return are public and permanent.

Monthly, an IC drift report recalibrates the Rank IC diagnostics. The point of all of this is that the site never gets to grade its own homework: the scoreboard is forward-looking, real, and uneditable.

• SEC filings — 10 years of as-filed fundamentals via SEC Company Facts (the ground truth for quality, forensics, and demonstrated growth).
• Yahoo Finance — prices, analyst Analyst estimates, and coverage.
• FRED — Fed macro series behind the Macro flags.

The whole pipeline re-runs daily via GitHub Actions; the IC drift report recalculates monthly. Paper ledgers persist append-only with transaction costs in bps.

A searchable index of all 106 defined terms. Click any term chip to open its definition — or click a term inline in any section above.

This site is a research tool. It is not financial advice, not a trading bot, and not a crystal ball. Factors work on average over years, not on every stock every month. The system's own track record is honest and often humbling by design. Past performance — including paper-traded performance — does not guarantee future results. Do your own research.

Back to the Factor Lab Cockpit