Research & Methodology

Ticker Inputs: Ticker fields accept symbols separated by spaces or commas (SPY AGG GLD or SPY, AGG, GLD). Portfolio, benchmark, risk-on/off, and custom factor fields all share this format. Leveraged exposure can be specified with custom-ticker notation (e.g., SPY?L=2, or SPY?L=2&FR=TBILL to use T-Bill financing).

Weight Inputs: Weight fields follow the same order as their tickers; fixed-weight portfolios and benchmarks must sum to 100%. When optimization is enabled, the optimizer determines portfolio weights. When tactical allocations are enabled, the standard portfolio inputs serve as the risk-on state as a new allocation will be required for the risk-off state.

Cash-flow Inputs: Positive cash flows are recurring contributions, negative flows are withdrawals, and the frequency dropdown schedules them.

Inflation adjustment: Inflation-adjusted views convert portfolio values, benchmark values, and return metrics into real purchasing-power terms using the CPI series in the ticker mapping table below.

Factor sources: Fama-French Factor data is sourced from the Kenneth French Data Library; Q-factor data is sourced from the Hou-Xue-Zhang Global-Q library, and updated monthly.

Custom factor tickers: User-selected factors allow for any ticker to be added to the selected factor model or to be used to construct a custom factor model. This lets an individual stock, ETF, or index return series serve as a factor.

Ticker mapping: Some funds have limited historical data. The following table is a mapping catalog to provide viable proxies or simulated data/tickers for some of the most widely recognized funds, indexes, and asset classes. Long-history simulated tickers splice published index, fund, yield, crypto, or composite proxy histories with live ETF/fund returns when available.

Backtest history limitation: A run begins at the latest first-available date among the portfolio, benchmark, and required proxy series. When a single ticker constrains the start date, Simfolio identifies that ticker beside the backtest inception date.

Live Performance: Provides a view of the portfolio's performance over user specified windows.

Actions: Previous and next scheduled optimization, rebalancing, or contribution/withdraw event for the modeled portfolio.

Live Portfolio Allocation: The current amount allocated to each holding. Allocations drift daily based on market fluctuations and are updated to reflect the current state of the modeled portfolio.

Transaction costs: Effective spreads are derived from each asset's daily OHLC using the Abdi-Ranaldo (2017) microstructure estimator and used to estimate slippage & transaction costs (.1% - .5% ∼ average) every one-way trade.

Walk-forward optimization: When optimization is enabled, the historical price data is divided into training and forward evaluation windows (with a purged execution gap between them - López de Prado, 2018). Weights are fit based on the desired optimization model using only the training window, then applied to later returns as the windows roll forward at the user's rebalance cadence.

Lookback window: Each optimization model's training window can roll (fixed length, slides forward) or expand (grows monotonically from a minimum sample). Rolling emphasizes recent regimes; expanding accumulates data over time.

Optimization models: Simfolio exposes 26 academically grounded optimization objectives: hierarchical allocation (HRP and HERC), risk budgeting, Tail Risk Budgeting (CVaR), Drawdown Risk Budgeting (CDaR), Inverse Volatility, Minimum Volatility, Minimum Downside Volatility, Minimum Mean Absolute Deviation, Minimum Tail Loss (CVaR), Distributionally Robust Tail Loss (CVaR), Minimum Worst Loss, Minimum Drawdown Risk (CDaR), Minimum Ulcer Index, Minimum Gini Mean Difference, Maximum Gini Ratio, Maximum Sharpe, Maximum Sortino, Maximum Omega, Maximum Calmar, Maximum Ulcer Performance Index, Maximum CVaR, Maximum CDaR, Maximum Diversification, Maximum Mean-Variance Utility, and Maximum Log Utility. The list covers the major peer-reviewed mean-variance, downside-risk, tail-risk, drawdown-risk, diversification, robust optimization, and Kelly-style utility families.

Stabilization: Covariance-sensitive objectives use Ledoit-Wolf shrinkage (2004); mean-sensitive objectives shrink the expected-return vector toward a Bayes-Stein grand mean (Jorion, 1986). Both reduce optimizer over-reactions to sample noise and improve walk-forward stability.

Constraints: Per-asset minimums, maximums, and group constraints bind the optimization model during fitting and constrain the portfolio from over/under allocating to assets or groups when optimizing.

Factor model specification: Factor regressions use OLS with Newey-West HAC standard errors. Best Fit compares the available academic factor specifications and displays the one with the strongest adjusted R²; Best Fit + Custom adds user-selected custom factor tickers to that comparison; Custom uses only the selected custom factors.

Selected production model: The live model is a portfolio-level stochastic-volatility forecast with a horizon-credibility mean rule, HAC drift-uncertainty controls, adaptive Metropolis MCMC, and filtered stationary-bootstrap shocks. It forecasts the portfolio return stream directly instead of forecasting every asset separately. In plain English, it estimates the slope of the road, reads the market weather, then drives many realistic future routes using bumps from the portfolio's own history.

Mean return layer: The mean is the gentle slope of the road, and it is the easiest part to overtrust. Simfolio starts with the portfolio's realized return evidence, measures drift uncertainty with HAC statistics, and lets the return signal shrink toward zero as the forecast gets farther out. That keeps one noisy historical average from steering every long-horizon scenario.

Stochastic volatility layer: Volatility is the weather on that road. The model treats risk as a hidden state that changes day by day instead of assuming one constant volatility number. Log-volatility follows an AR(1) process, meaning tomorrow's risk partly depends on today's risk, and a leverage effect lets negative return periods raise future volatility when the data supports it.

Adaptive Metropolis layer: MCMC is the part that asks, "which hidden-volatility settings could plausibly have produced this history?" The sampler tries many settings, keeps plausible ones more often, and uses those accepted draws to carry parameter uncertainty into the forecast. During warmup it learns a better proposal shape from the portfolio's own posterior draws, then freezes that proposal before retained samples are used.

Filtered stationary-bootstrap shock layer: Shocks are the bumps in the road after the weather has been removed. The model first divides returns by the inferred volatility state, creating filtered empirical standardized residuals that are closer to comparable day-to-day surprises. It then resamples those filtered shocks with a stationary bootstrap. The bootstrap draws random-length blocks using an automatic block-length rule, so short-term clustering and empirical fat tails can survive in the simulated paths instead of being replaced by an iid normal assumption.

How the pieces fit: The full process is: clean the portfolio return history, estimate a horizon-credibility mean, infer the hidden volatility path, sample plausible stochastic-volatility parameters, simulate future volatility paths, draw block-resampled filtered shocks, reapply the simulated volatility, and convert the result into forecast wealth paths. The tables and charts then summarize those paths as percentiles, drawdowns, loss probabilities, VaR, CVaR, and rolling metrics.

Out-of-sample validation: Before promotion, forecast candidates are tested with rolling-origin out-of-sample (OOS) validation: each model is fit only on information available at a historical date, then scored against what actually happened later. The research program tested more than 300 methodology variants, comparing forecast distributions against realized portfolio outcomes, a naive iid bootstrap benchmark, and prior production candidates across many portfolios, rolling origins, and horizons.We use adjusted CRPS internally because it scores the whole forecast distribution, not just one point estimate. The promoted live model had the best adjusted out-of-sample CRPS in the final promotion panel. Exact research scorecards are not published on this page; the public takeaway is that the live model must keep passing rolling-origin OOS checks before it stays promoted.

Long-horizon example: The long-horizon validation focus is not whether the median looks clean; it is whether the full downside, middle, and upside range is realistically sized when uncertainty compounds. For a retirement or withdrawal plan, that can be the difference between a forecast that looks deceptively smooth and one that better reflects bad-decade risk.

How we scored: The main score was adjusted CRPS (continuous ranked probability score), which compares the full forecast distribution with what actually happened. CRPS is not a percent accuracy number; lower is better. It rewards forecasts that put enough probability near the realized outcome without making the range unrealistically narrow or uselessly wide.

Benchmark comparison: The naive iid (independent and identically distributed) bootstrap benchmark is a simple Monte Carlo engine: it resamples daily shocks as if sequence, volatility clustering, and market regimes did not matter. The production model's advantage comes from keeping empirical shocks and tails while also modeling changing volatility, which becomes especially important when forecast uncertainty compounds across long horizons.

Reproducibility: Every run uses deterministic random seeds, so the same inputs, horizon, and simulation count produce repeatable forecast paths.

Simulation count: Current production runs use a fixed simulation budget so forecasts stay comparable from one portfolio to the next.

Forecast tables: Forecast tables keep the median as the fixed reference column and let the comparison column switch among percentile outcomes, mean, maximum, minimum, backtest values, VaR, and CVaR (conditional value at risk). CVaR summarizes the average outcome inside a selected tail, so CVaR 5% is the average of the worst 5% of simulated terminal outcomes.

Rolling metric paths: Forecast rolling metrics use one actual simulated path selected by terminal wealth outcome, then compute the selected rolling metric on that path. This keeps the displayed path internally consistent instead of stitching together unrelated percentile points from different simulations.

Drawdown paths: The forecast drawdown chart ranks each simulated path by its own worst peak-to-trough decline over the forecast horizon. Minimum shows the path with the worst drawdown, Maximum shows the least severe drawdown path, and percentile choices move from worse to better drawdown outcomes.

Strategy library: A catalog including time-tested buy-and-hold portfolios, tactical portfolios, optimized portfolios, and hybrids that combine both.

Filter and sort: Strategies can be filtered by metric, or whether they are tactical, optimized, both, or neither.

Classification: Trend, momentum, oscillators, drawdown, volatility, and market-state signals — including moving-averages, rate of change, RSI, Bollinger bands, realized volatility, VIX, cross-sectional relative momentum (Jegadeesh & Titman, 1993), and dual-momentum rules (Brock, Lakonishok & LeBaron, 1992; Moskowitz, Ooi & Pedersen, 2012).

Application: Signals can be applied at the asset level (per-holding overlay) or at the portfolio level (regime filter), letting users test both tactical sleeves and broader risk-on/risk-off regimes.

Combinations: Multiple signals can be combined with AND or OR rules.

Key Performance Metrics: A consolidated table highlighting key metrics that can quickly help users determine if a portoflio is worth further analysis.

All Metrics: The All Metrics dropdown is the full metrics suite. Backtests contain return path, drawdown, risk-adjusted, tail, gain-loss, distribution, and withdrawal-rate rows into Performance, Risk & Distribution, followed by benchmark comparison, statistical validity, time-series diagnostics, attribution and implementation, factor model, and Principal Component Analysis. Forecasts show model specification first, then forecast path metrics; terminal scenario probabilities and distribution columns stay in their dedicated forecast tables.

Rolling charts: Backtest rolling metrics recompute the selected metric over trailing one-year windows. Forecast rolling metrics use the selected single simulated path, chosen by terminal wealth target, so users can inspect how downside, median, upside, mean, minimum, maximum, or CVaR-like terminal outcomes behave through time.

Historical stress tests and efficient frontier: Historical stress tests are synthetic stress estimates based on the portfolio beta versus the S&P 500. The efficient frontier plots feasible risk-return combinations from the selected asset universe and marks the portfolio, benchmark, and tangency portfolio.

Explore sorting metrics: Explore sorting includes every portfolio-level backtest metric that can be compared across portfolios.

Returns and variability: Returns, standard deviation, downside deviation, drawdown, and gain-loss groups describe the pace, dispersion, and path of performance.

Tail risk: Value at Risk and Conditional Value-at-Risk (Rockafellar & Uryasev, 2000), maximum-drawdown analytics (Magdon-Ismail & Atiya, 2004), conditional drawdown (Chekhlov, Uryasev & Zabarankin, 2005), Ulcer Index and Ulcer Performance Index (Martin & McCann, 1989), Calmar (Young, 1991), Rachev (Biglova, Ortobelli, Rachev & Stoyanov, 2004), skewness, kurtosis, and the Jarque-Bera test (Jarque & Bera, 1980) focus on asymmetry and loss severity rather than average outcomes alone.

Capture and benchmark: Treynor (1965), Information Ratio (Goodwin, 1998; Grinold & Kahn, 2000), tracking error, up- and down-capture (Cumby & Modest, 1987), robust active return inference using Newey-West HAC standard errors (1987), bootstrap Sharpe-difference inference (Ledoit & Wolf, 2008), and factor alpha test whether headline performance survives relative comparison and aligned-history measurement.

Statistical validity: Newey-West HAC inference (1987), Lo-adjusted Sharpe (2002), Probabilistic Sharpe and Minimum Track Record Length (Bailey & López de Prado, 2012), Deflated Sharpe (Bailey & López de Prado, 2014), and stationary-bootstrap Sharpe-difference inference (Ledoit & Wolf, 2008) test whether reported performance survives sampling uncertainty and selection bias.

Attribution and factor diagnostics: Holdings attribution, Herfindahl-Hirschman concentration (Hirschman, 1964), risk contribution (Maillard, Roncalli & Teiletche, 2010), selected factor-model alpha, Fama-French factor exposures (Fama & French, 1993, 2015), regression diagnostics including Durbin-Watson (1950, 1951), Ljung-Box (1978), Breusch-Pagan (1979), and Brown-Durbin-Evans CUSUM (1975), loadings, and return attribution can help explain where realized returns likely came from.

Principal component analysis: Principal Component Analysis is computed from the realized asset-return correlation matrix in backtests. Eigenvalues determine variance explained, eigenvectors determine asset loadings, KMO and Bartlett diagnostics test whether the correlation matrix is suitable for dimension reduction, and Kaiser, broken-stick, Marchenko-Pastur, and parallel-analysis counts estimate how many components look meaningful rather than noisy. Portfolio PC risk share maps portfolio weights through the PCA covariance structure to show whether portfolio risk is dominated by one or a few latent common factors.

Implementation diagnostics: Smoothing and unsmoothing (Geltner, 1991; Getmansky, Lo & Makarov, 2004), non-synchronous and lagged-beta corrections (Scholes & Williams, 1977; Dimson, 1979), Treynor-Mazuy (1966) and Henriksson-Merton (1981) timing tests, and turnover groups catch the failure modes that simple CAGR-and-Sharpe reporting tends to hide.

Not investment advice: Simfolio is not a registered investment adviser, broker-dealer, or fiduciary. Nothing produced by the platform is a recommendation to buy, sell, or hold any security. Decisions and their consequences belong to the user.

Past performance: Every backtest is a replay of one realized history and cannot be assumed to always represent the future.

Costs: Slippage is estimated via the Abdi-Ranaldo (2017) microstructure estimator on daily OHLC. This is an estimate of effective half-spreads averaged across the observed sample and is not any broker's actual fill data, does not vary with order size, and does not include market impact for large orders. Taxes and margin interest are not reflected in any of the backtested or forecasted portfolio performance.

Walk-forward optimization can be suboptimal for the future: A model that fits and performs well across in-sample periods (historically) can fail in any future period.

Optimizer sensitivity: optimization objectives are highly sensitive to small changes in expected-return, covariance, and constraint inputs. Ledoit-Wolf shrinkage and mean shrinkage can reduce — but not eliminate this sensitivity. Two reasonable users running the same data with slightly different lookback windows, optimization frequencies, or constraints can experience vastly different results.

Forecast stationarity assumption: The forecast model is portfolio-level: it does not fit separate asset processes or an explicit regime-state model. The forecasting methodology is dependent on, and inferred only from the available portfolio history and proxy data. If the future market environment is vastly different, and outside anything represented in historical data, the forecast distribution will miss it.

Forecast drift uncertainty: Expected return is estimated from one realized return history, and mean estimation error dominates long-horizon portfolio forecasts. The horizon-credibility mean rule lets uncertain drift shrink toward zero as the forecast gets farther out, but the reported percentiles are still conditional on the observed sample.

Forecast horizon and history: Forward horizons that exceed the available history extrapolate stochastic-volatility persistence, residual blocks, and empirical tails beyond what was directly observed. A short sample can produce a long forecast mechanically, but market environments and tail events the sample never captured cannot be learned from that portfolio alone.

Tactical signal data-mining risk: Trend, momentum, oscillators, volatility, and regime signals are widely studied and widely retrofitted to favorable windows. Simfolio applies them to the user's data without look-ahead bias, but the choice of which signals, parameters, and combinations are exposed to multiple-testing bias and overfitting.

Live view is informational: The Live tab summarizes a model portfolio's current allocation and value.

AI summaries are not analysis: Portfolio News and AI Analyst outputs are language-model summaries. They can paraphrase, miscount, hallucinate, and omit context. They are to be used as a starting point for research, not as a verified second opinion, and never as advice.

Does Simfolio give investment advice?: No. Simfolio is a research tool for testing portfolio ideas, not an advisor or asset manager. Any investment decision remains yours, and you should consult a qualified professional for individualized advice.

What price data sources does Simfolio use?: Sourced from an institutional-grade data provider. We use adjusted close prices for all price data.

How accurate are long-horizon forecasts?: Long-horizon forecasts should be read as distributional ranges, not specific predictions. Simfolio reports percentile bands, loss probabilities, VaR, CVaR 1%, and CVaR 5% so users can see how wide the uncertainty is.

Can I import my existing portfolio from a broker?: Not yet. Enter tickers and weights directly, then hit the save button so that portfolio data will load upon login.

Do you offer refunds?: Yes. Paid plans include a 30-day money-back guarantee.

What happens if I cancel?: Cancellation stops renewal. Access remains active until the end of the current billing period.

What happens if I downgrade with too many saved portfolios?: Existing saved data is not deleted automatically. Creating additional saved portfolios is blocked until the account is back under the active plan's limit or the plan is upgraded again.