Research
Research Methodology
BestFolio aggregates published tactical and fixed asset allocation strategies and backtests them under a single, transparent framework. This page is the technical reference for everything you see on a strategy page: which strategies we accept into the catalog, how their backtests are constructed, how transaction costs and regime stress are modeled, how we validate robustness, and how we decide when not to endorse a strategy.
Our research and code are developed with the help of AI tools and reviewed before publication.
New to all this? Start with the plain-English map: where should your money go? It walks through the foundations and the main routes before you dig into the methodology here.
1. Inclusion Criteria
Not every strategy we encounter makes it into the catalog. Before a strategy is implemented, it has to clear four gates. Strategies that fail at any gate are either set aside or, if we implement them anyway for comparison, published in our rejection log with candid reasoning.
- Published, documented rules. The strategy must come from an academic paper, a book, a practitioner whitepaper, or a well-documented public post. We do not implement “secret” or proprietary strategies we cannot reproduce from public information.
- Mechanical, unambiguous signal definition. The rules must be specified tightly enough that two independent implementers would produce the same signal on the same data. Strategies requiring discretionary judgment or human interpretation are rejected.
- Implementable with publicly tradable instruments. Every ticker used must be investable by a retail account — ETFs, mutual funds, or index funds. Strategies that rely on institutional-only instruments, OTC derivatives, or unlisted alternatives are excluded.
- Minimum 10-year backtest window. Either the underlying ETFs have enough history, or our documented proxy chain can extend the series to a statistically meaningful window. Strategies that cannot be validated over at least one full business cycle are not admitted.
Note on editorial control. Passing the inclusion gates does not equal endorsement. Several strategies in the catalog — including the Kelly Signal family — are flagged as educational-only and documented in our rejection log with the reasons we do not recommend running them with real capital.
2. Data Sources
- Price data: Sourced from an institutional-grade market data provider, using adjusted close prices that account for stock splits and dividend distributions.
- Total return: Fund adjusted prices include distributions as the data provider records them, with the corrections documented in Section 6. Some historical index proxies are price indices with an estimated dividend yield added, or monthly total-return series, so those segments approximate total return and do not reproduce each historical ex-dividend date.
- Macro data: Select strategies use FRED series (unemployment, CPI, yield curves, corporate credit spreads) as signal inputs. Each macro signal applies its documented publication lag. First-release (vintage) values are used where available; some older observations and fallback sources use revised values.
- Frequency: Histories combine adjusted fund prices, index observations and modelled proxy returns, and their native frequency varies by source. Some older histories are monthly and are interpolated to the trading days of the New York Stock Exchange, so US market holidays never appear as trading days; those interpolated days are not measured prices. Daily history from markets that keep other days (the Frankfurt gold fixing, gold futures, the Nikkei 225, the dollar index, Treasury yields) is put on the same New York trading days: a price from a day the exchange was closed counts toward the next trading day's return. Signal and rebalance frequency are set per strategy (Section 3). Volatility counts each month in which most of the portfolio rests on interpolated data as one month-end return (see Volatility). A monthly strategy that weights its assets by volatility or picks them by low correlation measures every asset it compares on its last 12 month-end returns wherever the window of one of them rests mostly on interpolated days, because interpolated days barely move from one to the next and would read as an asset with no risk.
- Refresh cycle: Price data is refreshed daily. Signals are recomputed at the end of each month.
Credits and notices
Prices of US stocks, ETFs and mutual funds: Data sourced by Tiingo.
This product uses the FRED® API but is not endorsed or certified by the Federal Reserve Bank of St. Louis.
Series retrieved from FRED, Federal Reserve Bank of St. Louis:
- Source: U.S. Bureau of Labor Statistics via FRED: the unemployment rate (UNRATE) and US consumer prices (CPIAUCSL, CPIAUCNS).
- Source: Board of Governors of the Federal Reserve System via FRED: the federal funds rate (DFF, FEDFUNDS), Treasury yields (DGS1, DGS2, DGS10, GS1, GS10) and the dollar-euro exchange rate (DEXUSEU).
- Source: Federal Reserve Bank of St. Louis via FRED: breakeven inflation (T5YIE, T5YIFR).
- Source: Moody's via FRED: the Baa corporate bond yield (DBAA, BAA), used as a signal input. © 2017, Moody's Corporation, Moody's Investors Service, Inc., Moody's Analytics, Inc. and/or their licensors and affiliates (collectively, "Moody's"). All rights reserved.
- Source: Chicago Board Options Exchange via FRED: the VIX (VIXCLS), used as a signal input. Copyright, 2016, Chicago Board Options Exchange, Inc. Reprinted with permission.
- Source: OECD via FRED: the euro exchange rate before 1999 (CCUSMA02EZM618N), national consumer price indices (DEUCPIALLMINMEI, FRACPIALLMINMEI, ITACPIALLMINMEI, ESPCPIALLMINMEI) and, when the OECD's own service is down, its composite leading indicators.
- Source: Eurostat via FRED: euro-area consumer prices (CP0000EZ19M086NEST). Copyright, European Union.
Other sources:
- OECD (2026), Composite leading indicators, OECD Data Explorer, https://data-explorer.oecd.org/ (accessed on the date of each data refresh).
- OECD (2026), Main Economic Indicators, OECD Data Explorer, https://data-explorer.oecd.org/ (accessed on the date of each data refresh).
- Fama-French factors and portfolios: Ken French Data Library, Tuck School of Business at Dartmouth.
- Gold price, June 1968 to December 1998: Source: Deutsche Bundesbank (Frankfurt Stock Exchange gold fixing, converted to US dollars per ounce at the official US dollar fixing).
- Gold price, other months: World Bank Commodity Price Data (The Pink Sheet), licensed under CC BY 4.0.
S&P 500 is a registered trademark of S&P Dow Jones Indices LLC or its affiliates, and Russell 2000 is a trademark of FTSE Russell. Other index names are trademarks of their owners. BestFolio is not affiliated with or endorsed by them.
What you may do with third-party data is set out in section 3.5 of the Terms of Service.
3. Backtest Mechanics
- Rebalancing: Monthly, on the last trading day of each calendar month (unless the strategy’s original rule specifies a different frequency, in which case we preserve it).
- Signal timing: Signals are computed using end-of-month closing prices. The resulting allocation is applied to the following month. A signal dated Feb 28 uses complete February data and determines what you hold during March.
- Between signals: The backtest buys the target weights at each signal and holds those positions until the next one, so the weights drift with prices, as they do for a follower who only trades when a signal arrives. A monthly signal trades the drifted positions back to its target, even when the target has not changed, and the cost of that trade is measured from the drifted positions (Section 4). Daily strategies only send a signal when their allocation changes. Portfolios hold their strategies the same way between their rebalances: a fixed-weight portfolio moves back to its target weights at each month end, a walk-forward portfolio at each of its own rebalances.
- Look-ahead controls: Signals use only data dated on or before the signal date, and macro data is lagged by its publication date. Macro inputs such as the unemployment rate use first-release values from the FRED ALFRED vintage archive; where the archive is shorter than the backtest, only the earlier segment falls back to revised history, and each signal records which vintage it used. Where an older history exists only as monthly prices, the days in between are filled in for the portfolio's holdings, but a daily or weekly signal reads the last real month-end price instead, so it never sees a value filled in toward a month-end that had not happened yet. It also cannot react inside such a month, so daily-signal results over those years are approximations. The strategy card and page label those years "research reconstruction" and the growth chart shades them; the start date is kept.
- Start date: Each strategy starts at the earliest date that all its required investment and signal inputs support, after its lookback period. Where a strategy that decides at month-end shows results only from a set date, its indicators may still read the data before that date, so the first month shown already carries a signal with its full lookback. Daily and weekly strategies do not, because older data can be monthly prices filled in between month-ends; where their signal reads such an asset, it uses the last real month-end price, and the strategy page says which asset and when.
- Starting NAV: Normalized to 100 at the beginning of each backtest. The growth charts draw it from $10,000 (or €10,000), the starting amount Portfolio Visualizer uses, so the figures compare directly; every return and ratio is the same either way.
4. Transaction Costs & Stress-Adjusted Slippage
All backtests include realistic transaction cost modeling. Costs are not set to zero — they are deducted from portfolio returns at every rebalance.
- Base cost: 10 basis points (0.10%) per one-way trade, applied proportionally to portfolio turnover at each rebalance. A full 100% portfolio rotation incurs 0.10% in costs; a partial 30% rotation incurs 0.03%.
- Stress-adjusted slippage: During periods of elevated volatility, spreads widen and execution quality degrades. Our engine dynamically scales transaction costs by up to 3× the base rate when recent 20-day realized volatility exceeds the baseline (~16% annualized). Backtests already account for the higher friction you would experience during market stress (e.g., March 2020, Q4 2018).
- Turnover tracking: Every backtest reports total transaction costs, annual turnover, and average trades per year in the Summary Statistics panel. Friction impact is directly auditable.
- How turnover is measured: one-way. On each trade date we add half the sum of the absolute changes in every holding's weight, then divide by the years covered, so 100% a year means the whole portfolio sold and bought back once a year. A strategy's full-history figure is measured on its drifted holdings at every rebalance and counts the first purchase. A portfolio's is measured on its combined ETF holdings: moves between strategies (the monthly return to target weights, walk-forward reweights) count, and two strategies trading the same fund in opposite directions net out. Because a portfolio blends its strategies' returns rather than each fund's price, the rebalancing of drift inside each strategy is added at that strategy's own long-run rate, shown separately as an estimate. Over a chosen date window, turnover counts only the trades inside the window, not the holding it starts with. A tax on every buy and every sell costs about 2 × turnover × the rate a year; set your rate in Settings to see it.
- Taxes: backtests are gross of tax. Every metric we publish (CAGR, Sharpe, Sortino, MaxDD, Calmar, turnover, etc.) is pre-tax. All 110+ strategies use the same gross-return engine, so cross-strategy comparisons are apples-to-apples regardless of when each was added. Jurisdictional after-tax estimates live in the US Tax Impact panel on each strategy detail page (US federal rates, holding-period aware), and in the TBSZ analysis for Hungarian users. Other jurisdictions are on the roadmap.
- Fund expenses and borrowing: Real ETF and fund prices already reflect fund expenses, so no separate expense ratio is deducted from them. Synthetic leveraged proxies, used before a leveraged ETF existed, deduct a modelled expense ratio plus a borrowing cost on the borrowed amount (Fed funds plus a spread).
- What is not included: Margin interest for an investor who applies leverage at the portfolio level in their own account, and behavioural factors (delayed rebalancing, panic selling). These vary by investor and brokerage.
5. Survivorship & Selection Bias
- Faithful implementation. We implement each strategy as described by its original author(s). No proprietary optimization or curve-fitting is applied on top of published rules.
- Selection transparency. The catalog contains well-known, publicly documented strategies. This is a form of selection bias — we are showing strategies that gained attention, which may correlate with strong historical performance. We try to offset it by publishing the rejection log, which documents strategies we implemented but do not endorse. The Robustness score (Section 11) attacks the same problem quantitatively: it deflates every strategy’s Sharpe for the number of strategy variants we have tested, including the ones we never released.
- Ticker survivorship. Our proxy fallback chain (Section 6) uses mutual funds and index series that predate modern ETFs. The fund chain is visible in the Price Build Log on every strategy page.
6. Synthetic / Proxy Data (Fallback Chain)
Many ETFs have limited history. To produce longer backtests, we use proxy data for periods before an ETF existed:
- Fallback chain: Each ticker has a documented hierarchy of proxies. When the primary ETF has no data for a given date, the engine walks down the chain until it finds a valid source. Price continuity is maintained by scaling at each join point.
- Transparency: The full fallback chain for every ticker in a strategy is visible on the strategy detail page under the “Price Build Log” section.
Key substitution chains (100+ tickers covered):
- S&P 500: SPY (1993) → VFINX (1987) → S&P 500 total return, daily (1976: the index's daily closes plus each month's dividends from the monthly series) → S&P 500 total-return monthly series (1900). Until 28 September 2026 VFINX filled 1976-1986 too, but as the price providers serve it that fund carries no distribution before 1980 and none of its capital-gains payments of 1980-1986, so those years trailed the S&P 500 by 2 to 9 points a year (1985: +22.7% against +31.7%).
- Nasdaq 100: QQQ (1999) → ^NDX index (1985, price plus a 0.6% dividend yield) → Fama-French large growth, daily (1926). Until 28 September 2026 a Vanguard growth fund (VWUSX) filled 1960-1985; its provider history is month-end prints with no dividends and none of its capital-gains payments, which read as one-day drops of 7% to 14% each September, so it was dropped.
- US Small Cap: IJR (2000) → NAESX (daily from 1985) → Fama-French small cap, daily (1980); IWM (2000) → ^RUT index (1987) → Fama-French small cap, daily (1926)
- US Mid Cap: IJH (2000) → VIMSX (1998) → ^MID index (1981, daily S&P MidCap 400 closes, a price index with an estimated 1.2% dividend yield added) → Fama-French mid cap, daily (1926, the average of the value-weighted NYSE size deciles 6 to 8), which covers only the years before 1981. The Fama-French tier is a reconstruction of mid-cap behaviour, not the S&P MidCap 400 itself; over the 1998-2026 overlap it tracks IJH at 0.98 monthly correlation with a 3.7% annual tracking error.
- US REITs: VNQ (2004) → VGSIX (1996) → FRESX (1986) → NAREIT monthly (1971)
- Mortgage REITs: REM (2007) → FTSE Nareit Mortgage REITs monthly (1971)
- Int'l Developed: EFA (2001) → VGTSX (1996) → MSCI EAFE monthly (1969). Until October 2026 the September 1989 to April 1996 stretch came from an actively managed fund, T. Rowe Price International Stock (PRITX), which returned 9.3% a year there against 4.7% for the index, mostly by holding less Japan. The index now covers those years too. Before that fix, until September 2026, the same stretch pointed at PRTIX by mistake: T. Rowe Price's US Treasury index fund, so every developed-international ticker (EFA, VEA, VEU, VGK, IEFA, SCZ, VXUS) carried a Treasury fund's returns, +32% in 1990, the year MSCI EAFE lost 23%.
- Europe: VGK (2005) → VEURX (1990, the Investor shares of the same Vanguard fund) → European stocks monthly (1975, a reconstruction from the Ken French Data Library: 14 European countries, weighted to match French's Europe index) → MSCI EAFE monthly (before 1975 only). Until November 2026 every year before 2005 used MSCI EAFE, which was mostly Japan in the late 1980s (72% at the end of 1988), so “Europe” lost 24% in 1990 when European stocks were flat, and lost 1% in 1997 when they gained 24%.
- Japan: EWJ (1996) → T. Rowe Price Japan (1991), then Fidelity Japan (1992) where it has gaps, both actively managed → Nikkei 225 index (1965, in yen, translated to dollars at each day's exchange rate from 1971 and at the 360-yen Bretton Woods parity before, a price index plus an estimated 1.8% dividend yield). The Nikkei is price-weighted where EWJ is weighted by market value, so the years before 1991 have different sector weights.
- Int'l Developed, currency-hedged: DBEF (2011) → MSCI EAFE in dollars (the EFA chain above) multiplied by the US Dollar Index, which removes the dollar's move against a basket close to EAFE's currencies (1971). It leaves out the interest-rate carry a real hedge earns or pays.
- Emerging Markets: EEM (2003) → VEIEX (1994) → emerging-markets monthly series: the MSCI Emerging Markets index from 1988, reconstructed back to 1925 before that (no investable EM index existed before 1988, so that tail is research-only). Until October 2026 the November 1990 to April 1994 stretch came from an actively managed fund, Fidelity Emerging Markets (FEMKX), which returned 16.2% a year there against 32.6% for the index (1991: +6.8% against +59.9%). The index now covers those years.
- Long Treasuries: TLT (2002) → VUSTX (1986) → US 30-year Treasury monthly (1920)
- Intermediate Treasuries: IEF (2002) → VFITX (1991) → FRED 10-year yield, synthesized to total return (1962) → US 10-year Treasury monthly (1920)
- Short Treasuries: SHY (2002) → VFISX (1991) → FRED 2-year yield, synthesized to total return (1976) → T-bill monthly (1900)
- T-Bills: BIL (2007) → ^IRX yield (1960) → T-bill monthly (1900)
- TIPS: TIP (2003) → VIPSX (2000) → PRTNX (1998) → synthesized TIPS monthly series (1972). TIPS were first issued in 1997, so nothing before 1998 is a traded TIPS price: that tail is a research reconstruction, and the Price Build Log names it whenever it serves.
- Gold: GLD (2004-11-18) → GC=F futures (2000-08-30) → World Bank monthly averages, dated at each month-end (1999 to August 2000, a gap no free daily source covers) → Frankfurt gold fixing, daily (1968-06-18 to 1998-12-30: every fixing from the Deutsche Bundesbank, converted to dollars at the same day's official dollar fixing) → gold spot monthly (1920). Before August 1971 gold had an official dollar price, so it was not a free-floating asset.
- Commodities: DBC (2006) → ^SPGSCI (1984) → GSCI monthly (1969, energy-heavier than DBC's basket)
- Aggregate Bonds: AGG (2003) → VBMFX (1986) → US 10-year Treasury monthly (1920; rates only, no credit spread)
- High Yield: HYG (2007) → VWEHX (1980)
- Short Corporate Bonds: VCSH (2009) → VFSTX (1982, Vanguard's short-term investment-grade fund, actively managed but close to an index) → short Treasuries scaled 1.1x (the SHY chain above) before that
- International Bonds, currency-hedged: BNDX (2013) → PIMCO International Bond (U.S. Dollar-Hedged) (1992, actively managed: it beat BNDX by about 1 point a year where both exist, so its years flatter BNDX by about that much) → US aggregate bonds (BND, then the AGG chain above) before 1992, since a hedged foreign bond index moves with interest rates rather than currencies
- Mutual-fund tiers are total return. Where a provider's record of a fund's distributions starts late, the missing payments come from a second provider's record and are reinvested when the history is built: VWEHX, VWESX and VUSTX before December 1989, VBMFX before 1990, FRESX before March 1990, and a handful of early payments of the Fidelity Select, NAESX and Vanguard index funds. Until 28 September 2026 those years were price only, which left out 8 to 15 points a year of bond-fund income and 6 of REIT dividends.
- Managed Futures: KMLM (2020) → DBMF (May 2019) → Barclay BTOP50 monthly (a real index of the largest trend-following funds from 1987; the series before that is reconstructed). DBMF itself goes straight to the index before May 2019, and CTA uses the KMLM chain. Until October 2026 the 2008 to 2020 stretch used a Rydex managed-futures fund scaled 1.2x, which trailed the index by about 3 points a year.
- Sectors, small value and growth styles: Fama-French daily portfolios (1926) before each fund's own history: industry portfolios for seven sector funds (XLK, XLE, XLV, XLU, XLP, XLY and XLI), small value for VBR, IWN and similar, large growth for VUG, MGK and IWY. They are style and industry analogues, not the funds' own indexes. XLF and XLB have no Fama-French tier: they stop at Fidelity's actively managed Select Financials (1981) and Select Materials (1986) funds.
- Semiconductors: SMH (2000) and SOXX (2001) → PHLX Semiconductor index (1994, a price index plus an estimated 0.6% dividend yield) → Fidelity Select Semiconductors, an actively managed fund (1985)
- Leveraged ETFs: Synthetic daily-leveraged returns from underlying, with expense ratio and borrowing cost deductions
Several deep tiers are monthly series (institutional indexes and, for TIPS and emerging markets, research reconstructions), densified to daily and spliced like any other proxy. A strategy only reaches them when its own start date allows, and the Price Build Log on each strategy page names the tier that served every period, so you can see when a backtest is running on a reconstruction rather than a traded price.
Portfolio leverage before a fund existed
The Leverage setting on a portfolio moves part of each broad stock fund it holds into a 2x or 3x version of that fund, and its results show on the portfolio page and in the Leverage impact card. From a fund's first trading day, that swap follows the fund's own traded prices. Before a fund started trading (SPY in 1993, QQQ in 1999, IWM in 2000, EFA in 2001, EEM in 2003, USMV in 2011, and likewise for any other fund), leverage runs on the same stand-in price history the backtests use, the fallback chains above, as far back as that history goes, and the 2x or 3x version is modelled from it with its borrowing cost and fund fee.
A portfolio whose history starts earlier therefore shows the effect of leverage in the market falls before those dates too. Where the stand-in history has only monthly prices, it is turned into daily prices as in the backtests, so a 2x or 3x version built on it misses the daily volatility drag within those months.
7. Walk-Forward Validation
- Availability: Offered for strategies whose rules include tunable parameters.
- Method: A rolling-window approach — train on N months of history, test on the next M months, then slide the window forward and repeat.
- Purpose: Prevents overfitting to in-sample data by validating that a strategy’s edge persists out-of-sample.
- Identification: Strategies with walk-forward validation results are marked accordingly on their detail pages.
8. Walk-Forward Portfolio Optimization
Beyond validating a single strategy, BestFolio offers walk-forward optimization for multi-strategy portfolios. Given a basket of strategy variants, the engine rolls a historical window forward and re-optimizes the sleeve weights at each rebalance date using only data available at that point in time. The resulting track record is genuinely out-of-sample — the weights held in any given month were never fit on that month’s returns.
Available criteria. Seven objective functions can drive the optimization. Two of them implement classical Markowitz mean-variance optimization (Markowitz, 1952, Portfolio Selection) using mean and covariance estimates from the rolling window. The remaining criteria reflect alternative or post-modern frameworks that address known limitations of the mean-variance approach.
- Max Sharpe Ratio — The tangency portfolio on the Markowitz efficient frontier. Maximizes the ratio of expected return to portfolio volatility, with a cash reference of zero for now: the Sharpe shown across the site is measured above cash, and moving the optimizer to it is a separate change.
- Min Variance — The Global Minimum Variance (GMV) portfolio on the Markowitz efficient frontier. Minimizes portfolio variance without regard to expected return.
- Max CAGR — Directly maximizes realized compound annual growth rate over the lookback window. Closer in spirit to growth-optimal / Kelly sizing than to mean-variance.
- Min Max-Drawdown — Minimizes the worst peak-to-trough decline over the window. A path-dependent objective, distinct from variance-based risk.
- Max Sortino Ratio — Substitutes downside deviation for total volatility. A post-modern portfolio theory (PMPT) variant that does not penalize upside volatility. It measures the shortfall below zero for now: the Sortino shown across the site is measured above cash, and moving the optimizer to it is a separate change.
- Risk Parity — Inverse-volatility weighting (equal risk contribution). Developed explicitly as an alternative to mean-variance optimization, which is known to be sensitive to estimation error in expected returns.
- Equal Weight — A 1/N baseline. Often surprisingly competitive with optimization in out-of-sample studies (DeMiguel, Garlappi & Uppal, 2009).
Constraints & solver. All optimization is long-only with a user-configurable maximum weight per sleeve (default 40%) and weights summing to 100%. The problem is solved via sequential least-squares programming (SLSQP) at each rebalance date.
9. Parameter Sensitivity
Any strategy with tunable parameters (lookback windows, thresholds, top-N selection) or a choice of input series is at risk of being fragile: a tiny change to a parameter can flip a great backtest into a disaster. We address this at three levels.
- Defaults match the source. The default parameters published on each strategy page match the original author’s specification. We do not re-tune the defaults to make the headline numbers look better.
- Walk-forward covers weights, not parameters. Our walk-forward engine re-estimates sleeve weights on the trailing window at each rebalance date, so no weight is ever fit on data that includes the test period. It never re-tunes a strategy’s own lookbacks or thresholds, so it is not evidence about how robust those are.
- Sensitivity surfaces (roadmap). We are building a UI to display the full parameter sensitivity surface (e.g., CAGR and max drawdown as a function of lookback window in months) so readers can see how rigid or fragile a strategy’s published parameters are before committing capital.
- Input-series choice. A strategy that reads an economic series can depend more on which series it reads than on any lookback. Inflation Compass is the clearest case: over 2003–2026, with everything else pinned, its published breakeven gauge and a trailing-CPI gauge differ by about 6 points of CAGR and 22 points of maximum drawdown. Where this applies, the strategy page states the magnitude, not only the direction.
10. Regime Analysis
Aggregate metrics (CAGR, Sharpe, max drawdown over the full backtest) hide an important question: when did the strategy do well or badly? A tactical strategy that shines only in the 2009–2021 QE regime is not the same animal as one that held up in stagflation, the dot-com bust, and 2022 simultaneously.
- Full-period display. Every strategy page shows the equity curve and drawdown curve across the full available history, not a cherry-picked window. The backtest-period line on each page tells you exactly which years are in the sample.
- Stress windows. The Summary Statistics panel reports performance during well-known stress regimes (2000–2002 dot-com, 2007–2009 GFC, Q4 2018, Q1 2020 COVID, 2022 dual drawdown) where the data supports it. Strategies that survived only one of those regimes are clearly visible. The stress windows table puts every strategy and variant side by side through the same five windows.
- Regime-split metrics (roadmap). Our research pipeline includes a planned regime-tagging layer (growth/recession, expansion/contraction, high-vol/low-vol, rising/falling rates) to split every backtest into sub-period metrics. This is on the roadmap; ETA in a monthly release note.
11. Robustness Score (Deflated Sharpe Ratio)
Sections 7–10 stress each strategy on its own terms. The Robustness score answers a question aimed at us rather than at any single strategy: after backtesting an entire catalog, how well does this strategy’s Sharpe ratio hold up against the best that pure luck would have produced across everything we tried? The score is the Deflated Sharpe Ratio (DSR) of Bailey & López de Prado (2014), a number from 0 to 1. It is a historical statistical assessment, conditional on the trials we counted and on the moments estimated from each track record; it is not a probability that the strategy will keep working. It appears as the Robustness column on the leaderboard, as a Robustness badge on strategy pages, and as Deflated Sharpe in the Metrics Comparison table on the Performance page. All three surfaces share one computation, so the numbers always agree.
A catalog is a multiple-testing machine: run enough backtests and the best one will look brilliant by chance alone. A raw Sharpe ratio ignores that. The DSR deflates it with three corrections:
- Selection breadth. N is every strategy variant we have ever backtested and kept a result for, released or not and retired or not: the leaderboard rows, the reference families we keep as baselines (the single-asset 200-day trend variants, for example), the research strategies that never made it to the site, and variants we have since retired. A trial is a trial whether or not we shipped it or kept it, so N is larger than what the leaderboard shows and retiring a variant never lowers it; the live value, with the number of retained variants and of usable trials, is in the Robustness tooltips. From the spread of Sharpe ratios across those N trials we compute the Sharpe the best one would be expected to show by pure luck, and a strategy’s Sharpe has to clear that bar rather than zero. Because N grows as we test more, scores drift down when we add strategies, published or not. That is the metric working as intended, not a bug.
- Track-record length. The same Sharpe is worth more over 50 years of monthly returns than over 8. Shorter histories earn lower scores.
- Non-normality. The computation uses each backtest’s actual monthly skewness and kurtosis rather than a normal assumption. Negative skew and fat tails, exactly the return profile that makes a raw Sharpe overstate how dependable a strategy is, reduce the score.
The result is the estimated probability, on the backtest alone, that the strategy’s true Sharpe exceeds the expected best-of-N-by-luck benchmark. Mechanically, the DSR is the Probabilistic Sharpe Ratio evaluated at that deflated benchmark, computed on per-period (monthly) Sharpe as in the original formulation. Scores below 0.90 are flagged as fragile. Most long-history strategies in the catalog score close to 1.0, so the cutoff keeps the warning on the genuinely fragile minority instead of colouring the whole field. A fragile flag does not mean a strategy is broken; it means its track record cannot yet statistically separate skill from luck, given how many strategies we tried.
Two honest caveats. The score is a property of the full track record, so it does not change with the period filter selected on the leaderboard. And even this N is a floor: parameter variations we tried and discarded are not logged as separate trials, and no N we compute covers every strategy the industry has ever tried, so treat a high Robustness score as necessary evidence, not sufficient proof. Reference: Bailey, D. H. & López de Prado, M. (2014), “The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality”, The Journal of Portfolio Management, 40(5). Our returns are monthly and their lag-1 autocorrelation has a median of -0.03 across the catalog, so the autocorrelation-aware standard error of Lopez de Prado, Lipton and Zoonekynd (2026) moves no score by more than 0.02; the 2014 formula stands.
12. Sharpe Stability Ratio
The Sharpe Stability Ratio (SSR) measures how consistently a strategy’s rolling Sharpe stayed above zero. Two strategies can have the same full-history Sharpe even when one earned its returns steadily and the other depended on a few good stretches. The Stability column on the leaderboard makes that difference visible over each selected period. The badge on a strategy page uses that variant’s full history in USD.
We take simple monthly returns from the last NAV observation in each calendar month, in the selected currency, and compute a rolling 36-month Sharpe: the arithmetic mean divided by the sample standard deviation, both of the monthly returns above cash (the same cash benchmark as the Sharpe ratio), with no annualisation. We divide the mean of those rolling Sharpes by their long-run standard deviation, estimated with Newey-West, Bartlett (triangular) weights and bandwidth 35. Each autocovariance is divided by the total number of rolling windows. The benchmark Sharpe is zero. We require at least 120 monthly returns (ten years) and show n/a if any rolling deviation or the long-run variance is zero. Stability is recomputed for every leaderboard period.
The source paper lets a data-driven rule choose how many lags the long-run variance uses (15 in its 29-year hedge-fund sample) and reports that the ranking barely moves with that choice. We fix it at 35 lags, the full overlap of a 36-month window, so every strategy is measured with the same estimator. That makes our values lower than the paper’s method would give, so compare orderings across sources, not levels.
A value near 0 typically means the Sharpe swung between positive and negative stretches. Around 1 suggests steady risk-adjusted performance across the regimes covered. Values far above 1 deserve suspicion rather than applause: Fairfield Sentry, a Madoff feeder fund, scored 1.55 in the Portfolio Optimizer example linked below. A high score is a reason to examine the record, not proof that the returns are dependable.
The period matters. In our September 5, 2026 catalog check, the S&P 500 scored 0.30 since 2000, and 156 of the 169 variants with data over that window beat it. Starting in 2008, the S&P 500 scored 0.74 and beat two thirds of them. A window dominated by a long bull market can make buy-and-hold look especially steady. Compare the same period and currency, and try several start dates before drawing a conclusion.
Reference: Bajo Traver, M. and Rodriguez Dominguez, A., The Sharpe Stability Ratio: Temporal Consistency of Risk-Adjusted Performance (SSRN 6344658). Thanks to Roman Rubsamen of Portfolio Optimizer for the suggestion and for the explanation of the ratio that prompted this addition.
The catalog-wide study behind this column is published as a working paper by BestFolio Research: The Sharpe Stability Ratio in Tactical and Static Asset Allocation Backtests: Evidence from 173 Published Strategy Variants (DOI: 10.2139/ssrn.7441662). The paper covers 173 published variants and reports findings unfavourable to the statistic as well as favourable ones. The replication archive holds the code and result tables behind every number in the paper.
13. Metrics Computation
All metrics are computed from the monthly NAV series unless otherwise noted.
| Metric | Definition |
|---|---|
| CAGR | Compound Annual Growth Rate. Annualized geometric return derived from the full NAV series. |
| Sharpe Ratio | (Mean monthly return above cash × 12) / (Sample standard deviation of those excess returns × √12). Each month’s cash return spans the same two month-end dates as the strategy’s. Cash is US Treasury bills in USD (BIL since 2007, the 3-month T-bill rate before) and the euro overnight rate in EUR (€STR since October 2019, euro-area call money from 1999, German call money before). A strategy that sits in cash scores about zero. If the cash series is missing or more than 7 days stale at an end of the period, the Sharpe is shown as unavailable rather than computed against zero. Computed on month-end NAV in the displayed currency, and one definition on every page and every period, including the leaderboard windows, the comparison tables and portfolios. It is not CAGR divided by the volatility column, which is a daily-basis figure. |
| Sortino Ratio | Same numerator as Sharpe (mean monthly return above cash × 12) over the annualized downside deviation of those excess returns, so upside volatility is not penalized. Downside deviation is the root mean square shortfall below cash, measured across every month rather than only the months that trailed cash: months at or above cash contribute zero. Cash is the same series, over the same dates, as for the Sharpe. A track that never trailed cash has no Sortino rather than an infinite one, and a missing or stale cash series shows it as unavailable, never computed against zero. Same monthly basis and same definition on every page and period as the Sharpe beside it. |
| Volatility | Annualized volatility: the standard deviation of returns, scaled to a year. Strategy pages, portfolio backtests and the leaderboard's strategy rows use daily returns (times √252). Library and shared portfolio pages, the walk-forward page and model portfolio rows on the leaderboard's Full History use month-end returns (times √12). It is not the Sharpe ratio's denominator, which is always monthly. Where a backtest runs on the monthly series of Section 6, the days inside each month are filled in at a constant rate, so their daily returns are not real. Every month in which most of the portfolio rests on such a series therefore counts as one month-end return (times √12), and the daily returns of the other months as daily returns (times √252); the two are combined in proportion to the months each covers. A strategy with real daily prices throughout keeps its plain daily figure. The daily Win Rate counts real daily prices only, on the same rule. A portfolio applies it to the weights of its sleeves (each sleeve rebuilt or not by its own history, idle cash counting as real), and a benchmark to its fixed weights. |
| Max Drawdown | Largest peak-to-trough decline in NAV, as a percentage of the peak. Which NAV depends on the view: strategy cards, portfolio backtests and the leaderboard's Full History view in USD use daily NAV, while windowed and EUR leaderboard views, the library and shared portfolio pages use month-end NAVs, which miss troughs inside a month and so read shallower. Strategy pages show both, labelled monthly and daily. Over a stretch rebuilt from a monthly series, the daily NAV moves in a straight line between month-ends, so its troughs there are month-end troughs and the real intramonth low may have been deeper than shown. |
| Recovery | How long the worst drawdown took to regain its previous high: from the last day at that high before the fall to the first day back at or above it, shown in years. It is measured on the same NAV as Max Drawdown in each view, so depth and duration describe one episode. A strategy still below that high shows the time so far, marked as still underwater. Longest DD is a different figure: the longest time below a previous high in any episode, often a shallower one. |
| Calmar Ratio (MAR) | CAGR divided by the absolute value of max drawdown, both over the same period and on the same NAV as the Max Drawdown shown beside it. A path-dependent risk-adjusted measure. Over a strategy's whole record this is the MAR ratio (compound annual return since inception over the maximum drawdown since inception), so the leaderboard labels it MAR on Full History and Calmar on trailing and custom periods. The original Calmar ratio looks back 36 months only. |
| Keller Ratio | The return adjusted for drawdown that Wouter Keller and Jan Willem Keuning introduced in their 2017 paper Breadth Momentum and Vigilant Asset Allocation (VAA), later named after Keller: K = R × (1 − D / (1 − D)) when R ≥ 0 and D ≤ 50%, and K = 0 otherwise, where R is the CAGR and D the maximum drawdown as a positive fraction. D / (1 − D) is the gain that recovers a drawdown of D, so the cut grows faster than the drawdown and takes the whole return at 50%, which needs a 100% gain to recover. K reads like a return: 10% CAGR with a 20% drawdown gives 10% × (1 − 0.25) = 7.5%. The authors measure D on month-end values. The leaderboard uses the CAGR and Max Drawdown of the same row, the pair Calmar divides, so on Full History in USD, where that drawdown is measured on daily NAV, K can read slightly lower than on their convention. Keller's later papers also report a stricter K25 = R × (1 − 2D / (1 − 2D)), zero at a 25% drawdown; the leaderboard shows the original. |
| SWR | Safe Withdrawal Rate. The maximum constant, inflation-adjusted annual withdrawal rate that survives every rolling 30-year window in the backtest. Withdrawals are scaled by recorded inflation in the display currency. In USD that is US CPI-U (CPIAUCNS from 1913, CPIAUCSL from 1947), so century-scale backtests are tested against the 1929 deflation and the 1940s and 1970s inflation rather than a flat rate. In EUR it is euro-area HICP from December 1996, chained back to 1955 on a synthetic euro-area CPI: the OECD consumer price indices of Germany, France, Italy and Spain, combined with fixed weights taken from Eurostat's 1999 euro-area HICP country weights (41%, 25%, 23% and 11% after normalising over the four). A window that starts before its inflation series (1913 in USD, 1955 in EUR) assumes 3% a year instead. Before September 25, 2026, EU mode used HICP alone, which covers no full 30-year window, so every EUR rate rested on that 3% assumption. |
| PWR | Perpetual Withdrawal Rate. The maximum withdrawal rate that preserves real (inflation-adjusted) capital over every rolling 30-year window, on the same inflation series as SWR in each currency. |
| Ulcer Index | The root mean square of the drawdown below the running high, taken at every observation, so a decline weighs more the deeper it goes and the longer it lasts. Lower is better. Strategy and portfolio backtests measure it on daily NAV; the leaderboard and the library on month-end NAVs. Every page and CSV download shows it as a percentage, the unit of the drawdowns it averages (6.79%, for example). |
| UPI | Ulcer Performance Index. CAGR above cash (the CAGR minus the cash benchmark's CAGR over the same start and end dates, the same cash as the Sharpe ratio) divided by the Ulcer Index (drawdown depth and duration). Higher values indicate better risk-adjusted performance with emphasis on drawdown pain. |
| Robustness (DSR) | Deflated Sharpe Ratio. Probability (0 to 1) that the Sharpe reflects a real edge rather than the luckiest pick among everything we tested, corrected for track-record length, fat tails, and the number of variants tried. See Section 11. |
14. Strategy Variants: SmartStack & SmartLeverage
Many strategies on BestFolio expose more than one tab in the variant selector (Standard, Leveraged 2x, SmartStack, SmartLeverage, and so on). These overlays keep the strategy’s underlying signal logic unchanged and modify only the execution step, i.e. which tickers and weights the base allocation is translated into. Two of those overlays are proprietary: SmartStackand SmartLeverage.
SmartStack (return-stacking overlay)
SmartStack layers a gold + managed-futures diversifier on topof the base allocation rather than funding it by selling equities. The engine walks each position in the base strategy’s signal output and applies a priority ladder:
- Prefer a return-stacked ETF if one exists for the position (for example, LQD can be replaced by RSBT, which embeds bonds plus managed futures in a single instrument).
- Otherwise, substitute a 3x leveraged ETF at 1/3 weight (SPY to UPRO, QQQ to TQQQ, TLT to TMF). This preserves full economic exposure to the original asset while freeing 2/3 of capital.
- If no 3x product is available, substitute a 2x ETF at 1/2 weight (VNQ to URE, VEA to EFO). This frees 1/2 of capital.
- For assets without a suitable leveraged product (DBC, AGG, BND), keep the full weight and layer an equal-sized managed-futures (KMLM) sleeve on top. This piece is managed futures only, with no gold counterpart.
- Cash-like positions (BIL, SHV, SGOV) pass through unchanged.
The split between gold and managed futures is not a fixed 50/50. Only the capital freed by the leverage substitutions (steps 2 and 3) is split 50/50 between GLDM (gold) and KMLM (managed futures). The step-4 overlay adds KMLM on its own, with no matching gold. So whenever the base signal holds a commodity or aggregate-bond position, managed futures carries more weight than gold; when it holds none, the two come out close to 50/50. Because the base strategy rotates its holdings every month, the realized gold-to-managed-futures ratio shifts month to month, and gold never exceeds managed futures by construction. The overlay is funded only by the capital the substitutions free. When step 4 asks for more managed futures than that (any month holding DBC, AGG or BND), the overlay is scaled down to fit and the strategy legs are never touched, so the base strategy always keeps 100% of its own exposure. Until September 2026 the engine scaled every position instead, strategy legs included, which left the base strategy at 80% of its own exposure in about a quarter of all months; the five SmartStack variants were restated when that changed.
Worked example. Suppose the base signal holds four assets at 25% each: QQQ, IWM, VEA (equities) and DBC (commodities). Steps 2 and 3 convert the three equity sleeves into fractional leveraged positions (TQQQ 8.3%, TNA 8.3%, EFO 12.5%) and free roughly 46% of capital, which would split into about 23% GLDM and 23% KMLM. DBC stays at full weight (step 4) and asks for an equal 25% KMLM sleeve. The overlay would then need 71% of capital against the 46% actually free, so it is scaled by about 0.65 and lands at GLDM 15% and KMLM 31%, while the four strategy legs keep their full exposure. The 16-point gap between gold and managed futures is the commodity sleeve after the cap.
The net result: the investor keeps full exposure to the base strategy and adds a diversifying return stream, all inside a standard brokerage or UCITS account with no margin and no futures.
Capital weight is different from exposure. A 3x substitution frees two thirds of that sleeve, so gold and managed futures can become the majority of the account. For example, a 100% IEF defensive signal becomes roughly 33.3% TYD, 33.3% GLDM and 33.3% KMLM: about 1.67× estimated product exposure. A 100% BIL signal stays 100% BIL. The defensive label describes the base signal; it does not remove leveraged Treasury or alternatives risk. Daily leverage targets do not promise the same multiple over a longer holding period.
Each SmartStack variant shows the latest, median and maximum ETF capital weights and estimated product exposure from its full published backtest, plus how often defensive observations held a majority directly in gold and managed futures. Gold holdings can include a base-strategy position, so they are not labelled entirely as an added overlay. RSBT embeds managed futures as well as bonds; it counts as 2× product exposure and is shown separately from direct GLD/GLDM and KMLM holdings. These estimates include cash at 1× and do not measure the gross notional of derivatives inside the funds.
Why we built it. Classic diversification funds gold or managed futures by cutting equities, which in a strong equity decade turns diversification into a drag. Institutional investors solve this with futures-based portable alpha. SmartStack reproduces the effect using only retail-accessible leveraged and return-stacked ETFs.
SmartLeverage (fractional leverage overlay)
SmartLeverage is a leverage dial from 1.0x to 3.0x that amplifies the base strategy’s total exposure while keeping total notional at 100%. It does not naively swap every ticker for its leveraged equivalent. Instead, it multiplies each weight by the dial to get a target exposure and starts every position in its 1x fund, which would cost more than 100% of capital. A two-pass greedy algorithm then frees exactly the excess:
- Pass 1 (prefer 2x): move positions to their 2x ETF, largest target exposure first, until total notional is back at exactly 100%. The last position it touches is usually split between its 2x and 1x funds. 2x products have less volatility decay and lower expense ratios than 3x, so they are the preferred lever.
- Pass 2 (escalate to 3x): only if every 2x upgrade still leaves notional above 100%, upgrade positions from 2x to 3x in the same order. A maximum-leverage calculator sets the ceiling for any given allocation (every position at its highest available factor), and a dial set above it is capped there.
- Cash is never levered: cash-like positions in the base allocation (BIL, SHY, SHV, SGOV, TBIL) are scaled by the dial like any other weight but always stay in their 1x fund. Because the passes stop at exactly 100%, a fully invested portfolio never gets a cash sleeve, even at the ceiling.
Example: 1.5x on a 60/40 SPY/IEF portfolio targets SPY 90% and IEF 60% of exposure, which at 1x would cost 150% of capital, so 50% has to be freed. Moving all of SPY to SSO frees 45%. The last 5% comes from IEF: 10% of its exposure moves to UST, costing 5%, and the other 50% stays in IEF. The result is SSO 45%, UST 5%, IEF 50%: 1.5x exposure on exactly 100% of capital, with no cash and no 3x fund. At 1.2x a 50/50 portfolio needs only part of one position: SSO 20%, SPY 20%, IEF 60%.
Difference at a glance
- SmartStack changes what you hold: it adds a gold + managed-futures overlay on top of the base allocation.
- SmartLeverage changes how much you hold: it amplifies the base allocation’s total exposure without changing its asset mix.
- Both are execution-layer transforms. The underlying signal, rebalancing cadence, and research logic of the base strategy are untouched.
- Both are opt-in. Toggling between Standard and a Smart* variant on a strategy page simply swaps the execution overlay and re-runs the backtest. The Asset Universe panel updates to show the actual tickers the variant trades.
15. Research Pipeline & Changelog
BestFolio is a living research platform. We publish new research every month: new strategies and variants once they pass the inclusion criteria, and revisions to existing implementations whenever a source paper is updated or a bug is reported. There is no fixed number of new strategies per month. A candidate that fails review goes to the rejection log, not into the catalog.
Current catalog: 110 strategies, 88 tactical + 22 fixed, 6 free forever.
Recently added. Latest additions to the catalog (most recent first):
| Added | Strategy | Author |
|---|---|---|
| 2026-04-11 | Dynamic Macro Allocation (Sadek) | Bill Sadek |
| 2026-04-11 | Alpha-One Momentum | Bill Sadek |
| 2026-04-01 | Composite Momentum | BestFolio Research |
| 2026-04-01 | Growth-Inflation Sector Timing | Inspired by David Varadi (CSS Analytics) |
| 2026-03-27 | Golden Ratio | Bogleheads Community |
See the full strategy catalog for the complete list, filters, and sort order.
16. Rejection Log
Any research platform that only publishes strategies it likes is cherry-picking. We maintain a public rejection log documenting strategies we implemented honestly, backtested under this same framework, and then decided we could not recommend with real money. Each entry includes the strategy rules, our backtest metrics, the specific failure modes, and our verdict.
17. Behaviour Check (live vs backtest)
A backtest is a claim about how a strategy behaves. Once it runs live, that claim becomes checkable, but not on returns. At 15% annual volatility, telling a 2 percentage-point difference in annual return apart from noise takes roughly 225 years of live data. Ten months of live returns says nothing at all.
Behaviour is different. How often a strategy trades, how much of the book it moves, how often it sits in cash: these have a far smaller spread relative to their average than returns do, and implementation faults show up in them as a consistent offset rather than as noise. So behaviour answers in months what returns need decades to answer. This is what the Behaviour Check on each strategy page measures.
How a strategy is scored
We take the strategy's last six rebalances and compare them against every comparable six-rebalance stretch in its own history. Comparable means the market looked like today: windows are matched on the broad equity benchmark's trailing 12-month return and volatility, using fixed thresholds so nothing about the future leaks into the definition of a normal market.
Comparing whole windows rather than individual months is the point. A strategy sitting in cash for six months is not six independent coin flips, and scoring it that way overstates how unusual it is. When we tested the naive month-by-month version over 87 strategies and 176 months it flagged more than a third of the catalogue in a typical month, and the alerts in its lower, “unusual” tier typically lasted a single month. The window method flags about 3.5% of the strategies it can score in a typical month, and a flag has to hold for two rebalances before it is shown at all.
The same 15-year test shows the check's main limit. In about one strategy-month in five there was not enough comparable history to score it at all, because the market was in a state the strategy's own past had rarely seen. The check says so rather than guessing, and that gap is widest when markets are unusual: most of 2020, 2022 and 2023 fell into it. Of the months that could be scored, 95% read “in line”, a strategy scored two months running kept the same verdict 97% of the time, a typical flag lasted two months, and 21 of the 87 strategies were never flagged.
When our own history is wrong
A second, separate check re-runs each recently published signal through today's backtest engine and compares the result with the signal that is in your account. These are different questions: the first is about the strategy, this one is about us.
Some strategies read economic data that is revised after publication. When that happens, a backtest computed today can place the strategy somewhere different from where the live signal placed it at the time, purely because the underlying data has since changed. Where we detect this, the strategy carries a History under review notice naming how many rebalances are affected. In those months, the signals you received are the record; the backtest curve should be treated as provisional.
We publish this rather than quietly correcting it. A ticker that has merely been renamed to its canonical equivalent is not counted, and neither is a single isolated month: a notice appears only when the disagreement persists and changes the actual risk posture.
What it does not tell you
- It is not a return forecast. A flag predictsdifference from the backtest's typical behaviour, in either direction. It says nothing about whether a strategy is about to do well or badly.
- It does not prove the edge is real. It checks that the live version is doing what the tested version did, which is a narrower and more answerable question than whether the backtest was overfit in the first place.
- It cannot score every strategy. Coverage is currently 174 of 225 variants. Annual strategies do not rebalance often enough to fill a six-rebalance window and are not scored at all; quarterly and event-driven strategies are scored only once they have traded enough times. Any strategy also goes unscored while the market is in a state its own history has rarely seen, so the check has least to say when markets are least like the past.
18. Limitations & Disclaimers
- Hypothetical results: All backtested performance is hypothetical. It does not represent actual trading and was not achieved with real capital.
- No guarantee: Past performance — backtested or live — does not guarantee future results.
- Remaining frictions: Transaction costs and stress-adjusted slippage are modeled (Section 4). Fund expenses are reflected in real fund prices and deducted from synthetic leveraged proxies. Taxes, margin interest on leverage you apply in your own account, and behavioural factors (panic selling, delayed rebalancing) are not captured.
- Execution assumption: the backtest fills every rebalance at the closing price of the signal day (day D), so it earns the move from that close to the next. For daily strategies that read the same close they trade on, this is an execution assumption: a real order placed at that close would fill near it, not exactly at it. The schedule we publish for followers is the next session’s open (09:30 ET), which does not capture that move. This is a fill-timing gap, not look-ahead; the rebalance sensitivity card on each strategy page carries a “delayed close” line that trades one session later and bounds how much of the return rides on it. Next-open fills are not modeled: adjusted open prices are not in our price history.
- No universal winner: No strategy is guaranteed to outperform a simple buy-and-hold approach in all market environments. Tactical allocation involves active decisions that can underperform passive benchmarks for extended periods.
Questions about our methodology? Check the FAQ or reach out at [email protected].