Methodology

Data and universe

Return data is sourced from a third-party market-data provider as monthly total returns, adjusted for splits and distributions and resampled to month-end: dividend-adjusted net asset values for mutual funds, and dividend-adjusted closing market prices for exchange-traded funds (so an ETF’s figures match its issuer’s market-price returns, which can differ from its NAV returns by a few hundredths of a percent a month). The universe comprises the largest US-listed funds by reported assets, together with index proxies and a curated set of additional share classes that are force-included irrespective of reported size. It is refreshed nightly and currently contains 31,962 funds, with data through 2026-09.

Fund history begins at 1980-02. The month axis itself extends to 1950-02, but that earlier range is populated only by index series; no fund carries a return before 1980-02.

1. Total return, not price. Distributions are treated as reinvested. The S&P 500 total-return index is the default reference for US equity funds; the other equity indices offered as benchmarks are price-return series, labeled as such, and the exchange-traded proxies are total-return. Alpha and information ratio measured against a price-return index are biased upward by approximately the index dividend yield. Each fund page opens against the fund’s natural yardstick: US aggregate bonds for bond funds, MSCI EAFE for international equity, the Russell 2000 for small caps, Treasury bills for cash, and the S&P 500 otherwise.
2. Survivorship. The universe consists of funds currently in existence. Funds that have closed or merged are absent, which biases long-horizon category statistics upward. This limitation is disclosed rather than corrected.
3. Recency. The current, incomplete month is excluded. In addition, the most recent complete months are admitted only once their income has arrived: the provider folds distributions into the adjusted net asset value with a lag, so a recent month can carry the ex-date price decline without the distribution. A month in which the riskless 0–3 month Treasury bill funds deliver less than half of the bill yield is held back for every fund until the feed completes. Every profile states the period it covers.
4. Stable-NAV funds. A money-market fund holds its share price at $1.00 and pays its entire return as distributions, which the provider's adjusted series does not carry for these funds; the series is a flat 0.00%. Such funds are withheld from returns, risk statistics, screens and rankings, and their profile says so, rather than publishing a figure known to be false. They remain in the universe for fee and profile data.

Return and risk measures

Table 1 — Definitions of published return, risk and benchmark-relative measures.

Outlook score

The Outlook is a transparent quantitative score, not a rating issued by a ratings agency and not investment advice. Each fund is scored on within-category percentiles of four factors, which are combined into a weighted composite and mapped to five tiers (Strong, Positive, Neutral, Caution, Weak).

Table 2 — Outlook factors and weights.

Weights are re-normalized over the factors that are present. A fund is returned as Not rated where the expense ratio is unavailable, where history is shorter than 48 months, or where fewer than 5 category peers exist. The history threshold follows from the persistence factor, which requires two non-overlapping measurement periods; a shorter series cannot express persistence and is not scored on it. A separate gate prevents inexpensive but persistently lagging funds from reaching the upper tiers. Each fund profile publishes the factor percentiles underlying its own score.

Third-party star rating

Where a quantitative star rating is supplied by the data provider it is passed through without modification. Coverage across the full universe is sparse; an em dash indicates that no rating was supplied for that fund, not that the fund failed a screen. The measure is backward-looking and peer-relative, is issued by a third party, and is distinct from the Outlook score described in section 3.

Peer categories and ranks

Category rank is a fund’s total-return percentile within its peer category over the selected period, where 1 denotes the best. Only funds with data covering the full period are ranked, and ranking is performed against the whole universe so that a filtered view remains correctly ranked. Categories containing fewer than 5 ranked funds are aggregated to the asset-class level, and the basis applied is disclosed on the fund profile.

The data provider classifies a minority of the universe. For the remainder a category is taken from the provider's holdings-based classification where the fund's own returns are consistent with it, set to money market for a fund that behaves as one, or otherwise inferred from the fund name using a curated rule engine, so that those funds also receive a peer group. An inferred category is marked on the fund profile and never overwrites a stated one. A fund whose name yields no clear signal is left uncategorized and ranked against its asset class rather than assigned a category that cannot be supported. A stated category that the fund’s own return series contradicts — for example a low-volatility classification on a series exhibiting materially higher volatility — is rejected rather than repeated.

Returns-based style analysis

Where holdings are unavailable, style is estimated from returns by constrained regression of the fund’s series onto a palette of style proxies, with non-negative weights summing to one. The result describes how a fund has behaved, which may differ from what it holds. Where a holdings-derived style is available it is preferred and labeled as such.

Equity style weights

The equity style grid distributes a fund’s stock holdings across nine cells, three size bands by three value–growth bands, and reports the share of the fund’s classified stock weight in each cell. Each holding is placed by the characteristics of its own issuing company, so the grid shows where the stock exposure is concentrated rather than where the average holding lies: a fund split between mega-cap growth and small-cap value has an average near the center of the box and almost nothing in the center cell.

Table 3 — Construction of the equity style grid.

Holdings are as of the fund’s most recent portfolio report, which can lag the current portfolio by up to a quarter. A stock without enough valuation or growth inputs keeps its size band but is left out of the grid, and the classified share is stated beside it. Foreign-listed holdings are not classified, so international funds generally show the single-point box instead.

Holdings look-through

Portfolio look-through aggregates each constituent fund’s reported holdings, weighted by the fund’s weight in the portfolio, to give exposure at the security and sector level. The proportion of each fund that could be classified is reported; uncovered weight is shown as such rather than redistributed across the classified remainder.

Portfolio construction and projection

A portfolio is defined as a dated sequence of rebalances rather than a static weight vector. Between rebalances, weights drift with returns, and the allocation shown is the weight that actually earned each month’s return. Trade attribution compares the realized sequence against buy-and-hold and against annual rebalancing over the same period.

Projections are produced by block bootstrap of the portfolio’s own historical monthly returns. A withdrawal is escalated with inflation, preserving its purchasing power across the horizon; a contribution is not escalated, and is modeled at the amount entered. A projection can only resample the regimes contained in its sample, and the application discloses where the sample contains no material drawdown. Projections are hypothetical and are not a forecast.

Models: screening, blending, publishing

A model is a policy plus sleeves. The policy fixes the benchmark, the advisory fee, the main window (36 months by default, any length from 12 to 240 months), longer check windows, a tracking-error budget, a per-fund cap, a beta range, the stress episodes, the score formula and, optionally, a candidate list of tickers. Each sleeve has its own benchmark and screens its own candidates — the asset class and categories it names, or a pasted list (its own, else the model's) — and may be measured over a window of its own, typically starting when its youngest shortlisted fund starts. A sleeve's Screen, Compare and Blend use its window; Assemble, the stress test and the model's numbers use the main window.

Net of the advisory fee. The fee is billed a quarter at a time, so a fund's net monthly return is (1 + gross) × (1 − fee / 4) − 1 in a billing month (March, June, September and December in arrears; January, April, July and October in advance) and the gross return otherwise. Benchmarks are never charged. An uploaded series marked “net of fee” is not charged again.

Measures. Over a complete window of n months: annualized return Π(1 + r)12/n − 1; excess = annualized fund minus annualized benchmark; tracking error = sample standard deviation of monthly (fund − benchmark) × √12; information ratio = excess ÷ tracking error; IR/TE = IR ÷ (tracking error in percent); beta = slope of fund on benchmark; capture = annualized return in the benchmark's up (down) months, fund ÷ benchmark × 100; maximum drawdown measured from the growth of $1, the starting dollar counting as a peak; rolling excess (t3xR, t12xR) = fund growth minus benchmark growth over the trailing 3 or 12 months. The firm's second tracking error is measured on that 3-month excess: the sample standard deviation of the overlapping 3-month excess over each run of three months inside the window, × √4, with the information ratio and IR/TE on it defined the same way. A series missing any month of a window is reported as not covering it, never annualized over fewer months.

Benchmarks. A model or sleeve can be measured against an index or index fund, a fixed blend of them, inflation (CPI) plus a margin, or an uploaded monthly return series — the firm's own benchmark when the universe does not carry it. Benchmarks are never charged the advisory fee.

Screen score. Candidates are ranked on a weighted formula of measures, each net of the fee and against the sleeve benchmark. New models rank on net IR/TE alone (the information ratio divided by tracking error in percent, which favors steady outperformance with little deviation). Ready-made alternatives add low overlap (one minus the highest correlation of the fund's monthly excess returns with a fund already in the blend), consistency (the share of checks passed: each third of the main window and each longer check window in which the fund beat the benchmark), stress (average excess over the stress episodes covered), the median rolling 3-year information ratio over the last N years, the share of rolling 12-month periods that beat the benchmark, down capture, maximum drawdown, the expected IR and the expense ratio. A model can also build its own formula from up to eight measures — these, net excess, tracking error, up capture, the IR on 3-month excess and the length of record — each over the main window or the last N years where that applies, at any weights (scaled to add up to 100%). Each measure becomes a percentile rank, the number of ranked values below x ÷ (the count − 1) — in a spreadsheet COUNTIF(values, "<" & x) ÷ (COUNT(values) − 1), with ">" for a lower-is-better measure such as tracking error, down capture or expense — taken on the measures rounded to 12 significant digits so near-identical funds tie, a missing measure taking 0.5, and the score is their weighted sum × 100. Only funds that cover the whole window and fit the sleeve rank: R² against the sleeve benchmark must reach the sleeve's minimum (0.75 for US large cap by default; none for alternatives and inflation hedges, which are chosen for not tracking an index). Share classes of one fund are collapsed to the cheapest class that covers the window, and vehicles a brokerage account cannot hold (529 college-savings portfolios, insurance-only portfolios, collective trusts) are left out by default.

Evidence. An information ratio from a short record is mostly noise. The expected IR takes the IR over the longest of the last 10, 7, 5 or 3 years that the fund fully covers and shrinks it toward zero by that length: IR × years ÷ (years + 3). Records under three years are labeled thin. Style tilt regresses a fund's monthly excess on value, size, momentum and quality spreads built from index funds; when the spreads explain at least 40% of it, the dominant tilt is named, because that part of the edge can be bought with an index fund. The intercept, annualized, is reported as alpha after style.

Blending. A blend is a fixed mix rebalanced monthly. The optimizer maximizes the blend's information ratio exactly as it is shown — compounded net excess over the tracking error of the monthly net excess — subject to the cap, the tracking-error budget and the beta range, from several starting points. It aims 0.002% inside the budget so that rounding the weights to 0.1% cannot carry the tracking error over it, and the numbers shown with a suggestion are those of the rounded weights that will be saved. A limit counts as broken only beyond what the page shows: tracking error to 0.01%, beta to 0.01, a weight to 0.1%. When the budget and the beta range cannot both hold, beta is freed first; when even the budget cannot hold, the lowest-tracking-error mix is shown. Each fallback is stated. The robust mix averages the optimizer's answer with equal weight. A fund's contribution to excess is its weight × its average monthly excess × 12, scaled so the contributions add up to the blend's compounded net excess. The walk-forward check repeats the screen and the optimizer at each quarter-end from seven years to one year before the latest month, using only data available then, and records the blend's net excess over the following twelve months, so the edge a screen promises can be compared with the edge that survived. A month a held fund has no return for takes the sleeve benchmark's return after the fee and is listed. Funds that have closed or merged are not in the data, so they were never candidates: the check is, if anything, flattering.

The model and stress. The model's monthly return is the sleeve-weighted sum of the sleeve blends. Each sleeve's share of tracking error is its weight × the covariance of its excess with the model's excess ÷ the model's tracking variance. In each stress episode the model's cumulative return is compared with the benchmark's and with what beta alone predicts. When holdings did not exist during an episode, the replay is marked “with stand-ins”: each missing fund-month takes a stand-in's return after the fee — the sleeve's benchmark, or when that does not reach back the sleeve type's fallback fund, else the model benchmark — and every stand-in is listed.

Review and publishing. Each step records the inputs it was confirmed with; when an input changes, the step is marked for review with the reason. The inputs include the funds a blend holds (they set the screen's overlap measure) and, for an uploaded series, any upload that changed months already there or its fee treatment; a new month added at the end is new data, as it is for a fund. Publishing freezes the draft as a numbered version with a journal note and the numbers it had; client accounts follow the published version. Every change is recorded with the user who made it. A model or sleeve can be exported to an Excel workbook that recomputes its statistics, ranks, scores, consistency checks, blends, model and stress episodes with formulas from the monthly returns, and checks each formula against the application's value. The optimizer, the walk-forward check, style tilts and the long-run IR are searches or fits over other windows; the workbook reports them but does not rebuild them.

Client accounts: trades and tax estimates

A trade plan targets account value × (1 − cash target) × model weight for each fund in the published model's look-through, skips trades below the minimum trade size, sells in full any fund the model does not hold, and scales buys so cash never goes negative. Custodian exports are per position, not per lot, so the gain on a sale is estimated conservatively: losing lots are assumed to be sold first, and the estimate is the larger of “every reported loss realized, the rest of the sale at the winning lots' gain” and “only the sale's pro-rata share of the reported losses” — the worst case whatever the losing lots are worth. Gains are taxed at the long-term rate only when the newest purchase is more than a year old. The switching hurdle — tax ÷ (sale × the account's payback years, 3 by default) — is the extra return a year a replacement must earn to repay the tax; above 1% a year a sale is marked “consider holding”. Estimates, not tax advice.

Limitations

1. Monthly granularity understates intra-month drawdowns and daily-path risk.
2. Sharpe subtracts a Treasury bill rate; Sortino applies a minimum acceptable return of zero. The two are therefore not directly comparable. Alpha uses the same Treasury bill rate as Sharpe.
3. The universe contains only funds currently in existence, which biases long-horizon category statistics upward.
4. Expense ratios are unavailable for a substantial proportion of the universe. Filters on cost report the number of funds they could not evaluate; that figure should be read alongside the result.
5. All measures published here are quantitative heuristics intended for research. They are not investment advice and not a recommendation to buy or sell any security.

Each profile page states the exact period, benchmark and as-of date behind the figures it displays.