Quantitative Research

Portfolio lab

The published pages answer questions someone else asked. This one is the simulator behind them, opened up: choose the exit window, the holding period, how capital gets allocated when signals overlap, which slice of the universe to trade, what friction to assume, and how much money you actually have. Everything recomputes in your browser, on the raw events, every time you change something. Every published page on this site shows one configuration that someone else chose. This one recomputes the entire portfolio in your browser under whatever assumptions you set. Nothing here is precomputed except the raw events. The defaults are the pre-registered configuration — the one fixed before any result was seen. Everything you change from there is exploration, and the panel at the bottom tells you how much weight your exploration deserves.

Start from a preset

Four starting points, not four results. The published backtest reproduces the assumptions used on the rest of this site — flat 0.1% friction, unlimited positions. Measured friction swaps that flat number for the per-tier slippage measured from the data, everything else unchanged. Measured friction, 10-day hold also switches the exit window to 10 days, caps concurrent positions at 20, and turns on smart rotation. A $10,000 account, realistically starts from that same 10-day setup and adds what a small account actually faces: whole shares only, a $50 minimum position, and the position size capped at 1% of 20-day ADV.

Simulation setup

EntryWhere the simulated purchase happens. Filing day close is the default and what every other control on this page assumes — but it is only reachable for part of the sample: measured against EDGAR's acceptanceDateTime, just 33% of purchase filings are accepted during market hours. For the other 67% that close had already printed before the filing was public. Next day open is the honest version of "I read it tonight, I act at tomorrow's open". Real price at publication uses the actual traded price in the 1-minute bar containing the second of acceptance, and exits on a matching series from the same intraday provider, so no adjustment mismatch leaks into the return. Exit and smart rotation both work here; the short leg does not, because real prices were only collected for the buy side, and the next-open window is unavailable because only daily closes were collected. Careful comparing totals across the three: the real-price option only has data for ~4,400 events from 2019 onward, while the other two run the full sample since 2003. A smaller total there may just be less time, not a worse strategy. ExitHow long each position is held before it closes. Next open (window C) and next close (1 day) are the two tradable event-study windows used on the rest of this site — D is the default because closes are more liquid and the strategy does not need to know the exact hour a filing published. Every other value (2–30) holds for that many trading days instead. LegsThe short leg is built from a daily aggregate — how many insider sales happened each day and what they returned on average — not from individual events. Two consequences the page would rather state than hide. No universe filter touches it: sector, size, price, ADV and conviction all apply to the long leg only. And the measured per-tier slippage does not apply either, because the aggregate carries no price tier, so the short leg uses the flat value above. Borrow cost is charged per day held and matters: it runs from about 0.3% a year on easy-to-borrow names to well over 10% on hard-to-borrow small caps, which is exactly where this signal lives. Worth knowing before you read the result: sell events return −0.017% on average against +0.616% for buys, and show no gradient by intensity. The short leg is here as a falsification you can run yourself, not because it is expected to add anything. ExposureOnly applies when Legs is set to run both sides. Dollar-neutral splits the balance in half, one for longs and one for shorts, so gross exposure equals the starting balance. Full balance each side gives each leg the whole balance independently, which is 2x leverage in dollar terms. Position capOnly matters once Max concurrent positions is set below the default. Across both legs counts long and short positions together against that one number. Per leg gives each leg its own separate count, so a capped long book does not starve the short book of room. AllocationSix answers to the same question: where does the money for a new position come from. Sleeves is the published default — each day's budget is a fixed slice of the ORIGINAL balance, split evenly among that day's new signals, and never touches positions already open. Cash buffer caps each new position and simply skips it if the cash isn't there. The other four all target the same thing — equal weight across every open position plus the new ones — and differ only in whose money is used to get there when cash is short: Daily equal-weight rebalance trims every open position down to the new target size. Fund new buys pro-rata sells a slice of every open position proportional to its size. Fund new buys from winners sells the biggest gainers first. Fund new buys from losers cuts the biggest losers first and lets winners run. All four are always measured against the ORIGINAL balance, never the accumulated one — that distinction is what keeps a rotation-heavy configuration from silently compounding into a fantasy number. Balance ($)The capital the whole simulation starts with. Every allocation mode measures position sizes against this ORIGINAL number, never against the balance that has since grown or shrunk — a position opened today is sized off this number, not off today's equity. Max positionsCaps how many positions can be open at once. The default (999) is effectively no cap. Lower it and Position cap applies decides whether that count is shared between both legs or given to each separately. Smart rotationDo not sell and rebuy a name that re-signals while a position is already open. Whole shares

Execution & frictions

The published backtest assumes 0.1% per side. This strategy breaks even at 0.263% per side, and the spread estimated from our own daily highs and lows (Corwin-Schultz) sits above that in every price tier — read that as a ceiling, not as what you would pay. What we can measure directly is weaker than it first looked, in both directions. The 0.054% half-spread we measured in the small-cap third comes from a heavily selected subsample — names quoting continuously enough to build a full curve — and does not represent small caps. But the wider figures that appear when the filter is relaxed are not a measurement either: our intraday quotes come from IEX, and on names whose true spread is known to be 0.01-0.05% (over $200M average daily volume) the measured spread moves 13-fold with nothing but the number of quotes in the window — 0.596% where IEX showed fewer than 20, 0.045% where it showed over 500. Below roughly 500 quotes the number reflects feed coverage, not the spread. So the honest position is that we cannot yet measure execution cost where this strategy earns most of its return, and whether 0.1% per side is conservative or optimistic stays open until it is measured against a consolidated tape. That is why this is a control you can move, not a constant.

BrokerThe broker buttons only touch the friction fields, so you can combine any broker with any strategy above. They map to what each one actually charges on US equities: IBKR Lite is commission-free with smart routing, so the effective spread lands below the quoted one. IBKR Pro charges $0.005 a share with a $1 minimum per order — modelled here as a flat $1 per side, which is the term that bites on small positions and is the single most damaging cost at a small balance. eToro charges no commission but earns on the spread and gives no price improvement, so the effective spread is set to the full quoted estimate. Not modelled: eToro's ~0.5% currency conversion if you fund in anything other than dollars, and the fact that fractional shares often cannot be routed as market-on-close orders — this strategy trades at the close.

Slippage source Flat slippage (%) Eff./quoted (%) Short slippage (%) Borrow cost (%/yr) Commission ($) Commission (%) Min. position ($) ADV cap (%)

Universe filters

Three size axes, because they are not the same question. Price tier drives the bid-ask cost you actually pay and covers 100% of events — "penny" is the SEC definition, under $5 a share, which is a price, not a size. Market-cap bucket is the literature ladder and covers 74% (the rest have no share count filed before the event). Within-month tercile is size relative to the universe that month. Each filter shows what it costs you in events.

PriceBucketed by share price on the event day: under $1, $1–$5 (the SEC's own definition of a penny stock), $5–$10, $10–$50, over $50. Price, not size — a $2 stock and a $200 stock can sit in the same market-cap bucket.

Market capMarket capitalization at the time of the event, on the standard nano/micro/small/mid/large ladder. Needs shares outstanding filed before the event date — missing for about a quarter of events, which this filter then drops.

TercileSmall/mid/large thirds of the universe traded that same month — a moving ladder, not a fixed dollar one like market-cap bucket. Same shares-outstanding gap as the bucket filter.

SectorTwelve industry sectors, the same classification used on the Size & sector controls page.

RoleKeeps only events where at least one Officer, Director or 10% owner is among the signers that day. An event can match more than one role if several insiders filed the same day.

SessionWhich of the four IBKR-style windows the day's first Form 4 acceptance fell into, ET: pre-market (4:00-9:30), market hours (9:30-16:00), post-market (16:00-20:00), overnight (20:00-4:00). By this measure 36% of buy filings' first acceptance lands during market hours — a looser count than the Entry control's 33%, which additionally requires every filing that day, not just the first, to be in-hours. Either way most of the sample is outside market hours, which is where the filing-day-close anchor most often predates disclosure (see the Entry control's tip). This filter is about when the news came out, not the entry price, so it combines cleanly with any Entry option, including real price.

Conviction filters

ConvictionFour ways to measure how much an insider is betting, and none replaces another. Dollars bought covers 99.4% of events but confounds conviction with company size: a million dollars is nothing in a mega-cap and everything in a micro-cap. SNS is the paper's variable, shares bought over shares outstanding, neutral to size by construction — and the only one of the two that shows a gradient, with the top quintile returning 1.627% against 0.133% for the bottom. It covers 76.3%, and the quarter it drops is not random: those events carry a wider spread and a far weaker t-statistic. Distinct insiders is the cluster-buy measure most cited in the literature, though 87.5% of events here have a single signer, so it will discriminate little until the universe grows. Signer role is kept separate as a filter rather than a ranking, because roles are categorical: claiming a CFO outranks a director would invent an ordering that then decides results. Min. percentileType a percentile from 0 to 100, not a raw number — 80 means "keep only the top 20% by conviction". No more guessing whether 50000 or 0.000224 is a harsh cutoff or nothing at all; the scale problem disappears because you never touch the raw units. The percentile stays live: it is measured over the events that pass every OTHER filter you have set, so the actual cutoff moves on its own when you change sector, dates or size — typing 80 always means "the best fifth of what I'm looking at right now", not a number that quietly stops meaning that the moment you touch another filter. The reading below shows the raw threshold this resolves to and how many events it keeps, for reference. One caveat the buttons cannot fix: picking a percentile by looking at the whole sample is exploration, not a rule that existed beforehand — the out-of-sample split further down is what tells you how much that exploration cost you. Ranking windowWhole range (default) measures ONE fixed cutoff over the entire filtered sample. Day / Month instead measure a fresh cutoff for each calendar day or month, using the distribution of the PREVIOUS day-with-events or PREVIOUS calendar month — never the event's own period, which in production hasn't finished yet. This is the causal, laggable version of the ranking used to validate the live watcher's conviction filter, not a new conviction measure — it changes only WHEN the cutoff is measured, not WHAT is being measured. Min. price ($) Min. ADV ($) From To OOS cutThe cut splits the record in two. You are meant to choose your filters looking only at the calibration side; the out-of-sample side did not take part in that choice, so it is the only number that carries any weight. A configuration that looks brilliant before the cut and ordinary after it was fitted to the past, not discovered in it. Both sides are compared against SPY over their own dates, never against the full history. Exclude 2008 Exclude 2020 Hedge marketShorts SPY for the same notional as the long book, same day, while a position is open. Regime filterBlocks NEW positions only — open ones still exit on schedule. Frozen parameters (220-day EMA, 252-day volatility window, median threshold), not fit to this backtest's own out-of-sample window. Tested against SMA200, a golden/death cross, an HMM regime model, market breadth, and an industry-relative filter — none beat it out of sample. Known weak spot: sharp V-shaped corrections (checked against 2011), where it reacts late to the drop and then stays out of the rebound.

Result

Two benchmark lines, on purpose. Buy-and-hold SPY answers "would I have been better off just owning the index". Exposure-matched SPY answers "does the edge exist" — it only counts the days the book is actually invested, which for a one-day holding period is a fraction of the calendar. They can differ by an order of magnitude for the same index over the same years. Reading the wrong one is the easiest mistake to make on this page.

Drag on the chart to select a period.

Drag on the chart above, or either handle below, to change the date range without typing — both charts and every number on this page follow it.

Return by period, against both benchmarks

Same curve as above, cut into bars instead of a line — pick the bucket size. Every bucket compares your configuration, SPY buy-and-hold and IWM buy-and-hold over the exact same days, so a bar-to-bar read is fair.

Bucket size

Backtest composition

How the capital the simulation deployed splits across the fills in the selected date range — not a live position, there isn't one yet (see the note below). Total invested is the sum of entry notional across every fill shown; P&L in the tooltip is realized, not a mark-to-market of an open position.

What was verified before publishing this

The controls page publishes its sanity checks whether they pass or fail, and this one had been computing its own without showing them. The important one compares two independent paths to the same price: the multi-day series used here is built separately from the event windows the rest of the project uses, and if the two disagree on the day-after close, one of them is wrong and nothing below should be trusted.

The shape of the horizon

The selector above samples one holding period at a time. This runs all seventeen of them against whatever configuration you have set, so you see the relationship instead of a point. It matters because the two cases look nothing alike: with a conviction filter the Sharpe ratio falls monotonically from about 2.0 at one day, while without one it starts wiped out and then wanders flat around 0.7. Reading the whole curve is also the defence against picking a lucky number — an isolated peak on a noisy line is overfitting, a peak on a smooth one is just the shape.

Is it consistent, or is it one lucky stretch?

A filter combination picked after seeing the full-sample result and then judged on that same sample proves nothing. Persistent out-performance across independent stretches of time is a different claim, and it is a claim that can be checked. These numbers recompute with every change you make above.

How to read the trade count. A configuration that beats the index on 40 trades is noise with good luck. One that beats it across both halves of the sample, in most calendar years, on thousands of trades, and still clears the out-of-sample split, is making a much stronger claim. The specification counter tracks how many distinct configurations you have run this session, because the more you try, the more likely one of them looks good by chance alone.

Trade journal

Every fill the simulation makes, in MetaTrader Strategy Tester order — one row per round trip, entry and exit together, not one row per side. Entry and Exit are the split-adjusted series used to compute every return on this page, already including slippage; Gross (%) and Friction ($) pull that cost back out so you can see what the fill would have returned with no execution cost at all, and where the difference went.

What this still does not model

The entry price of the D window is not always reachable. Window D assumes you buy at the close of the filing day. But only 33% of purchase filings are accepted by EDGAR during market hours (9:30–16:00 ET): for the remaining 67% that close had already printed before the filing was public, so nobody could have bought there. Measured against real intraday execution prices (1-minute bars at the second of acceptance, exiting on a matching series from the same provider), the reachable entry costs about 0.08 percentage points per trade — around 6% of the portfolio result with no quality filter, and up to a quarter of it with one. The effect survives real execution prices in every configuration tested. An earlier version of this note put that cost near half the edge and said the effect did not survive without a filter; that came from measuring entry and exit on two different price sources with different dividend-adjustment conventions, and it does not hold once both legs come from the same one.

Read this before trusting any number above. Market impact is capped, not modelled: the ADV control shrinks a position rather than moving the price against you. Borrow availability is ignored, which matters if you ever enable the short leg. The universe is 788 SEC registrants that are still active today, so companies that went bankrupt or delisted are missing entirely — precisely where insider selling would be most informative. Corwin-Schultz estimates the quoted spread; the effective spread paid is typically 50-70% of that, and trading at the close is usually cheaper still, so the measured defaults are more likely too harsh than too kind. And every backtest on this page is a backtest: it is what the rules would have produced, not what you would have earned.