Help · App reference

Allocation models: the Backtest tab

From RangefinderInvest's built-in help · applies to version 0.59.3

Backtest is Allocation models' tearsheet, one of the three tabs (Allocation, Backtest, Optimize funds) sharing that workspace's model picker: select two or more models and see how each allocation would have performed. Read how the simulation works, then use the growth chart, drawdown track, monthly views, and a metrics table grouped into Return, Risk, Efficiency, and Consistency. Open it from the tab or from the Backtest row beneath Allocation models in the sidebar.

The comparison needs at least one model with a complete 100% allocation, judged the way Allocation models judges it: the slices add up to 100%, and each slice's funds add up to 100%. A slice with no funds at all (a Treasury ladder, a self-directed slice, or one you have not filled yet) has nothing to add up, so it counts as complete. The curve holds that slice at a steady value (no gain or loss) rather than spreading its share over the other funds, and the Warnings line names it. A model whose slices ALL have no funds has nothing to price, so it has no curve at all; the page says so and points to Allocation models, since downloading prices cannot change it. A slice you have not filled yet is named with its share of the model, whether it has no fund lines or only the empty 0% line the basic editor leaves. A fund line that holds a share of its slice but names no ticker is different: it makes the model incomplete, so the model is not listed until you name the fund or remove the line. If no model qualifies, the empty state distinguishes an empty library from incomplete saved models and offers Open Allocation models. Finish an allocation there, then return to Backtest; the tab remains available in Back history. Opening Backtest for one particular model that is not complete yet leaves it out of the comparison and says so, with a link to finish it.

Getting a fair comparison

  • Every curve is indexed from the latest common start date across the selected models, and runs to their earliest common end date, so no figure covers a period one of them lacks. The page names which model (and whether its price or weight history) pins the start. A model with young funds constrains everyone. Check the "Weights / Prices / Period" line under each column.
  • The chart, the table, the benchmark columns and the monthly views all cover that one window. When it starts on a day the market was closed (a version dated on a weekend, or a preset that lands on one), every model is measured from what it was worth that day, which is its last close before it.
  • Hiding a model with its legend chip only stops drawing that line. The window, and every number the table reports, follow the models you have selected. Untick one to widen the comparison.
  • A model's version history is replayed exactly as you recorded it. Each dated version applies from its effective date, and a diamond marks every switch with the weights that changed.
  • Extend back extrapolates early weights before a model's first version to lengthen the window. Its year options count back from each model's own first version; Max goes as far as the funds in that first version are priced. It is on by default at Max, and that stretch draws dashed; set it to Off to see recorded history only. Under a preset the From and To boxes show the window in force; a date you type stays as typed, and the caption under the chart names the dates actually used.
  • The vs benchmark (a ticker from your comparison list) draws as a dashed overlay and adds relative columns: excess CAGR, beta, tracking error, information ratio, up/down capture. To compare models, select them in the list on the left instead. Edit beside it opens the comparison-ticker list, the same shared editor the Overview's chart opens, so a ticker added here shows there too.

Three ways to read the curve

The toggle above the chart switches what the y-axis measures. All three describe the same backtest over the same window.

  • $: the growth of $10,000, the value the simulation actually tracks.
  • %: cumulative return from the window start, which is the fair comparison when models start at different values.
  • %/yr: annualized return so far, the per-year rate that compounds to the return at each date. It is the same figure the CAGR row reports, drawn as it settled, so a model that won on one early run shows its lead decaying rather than looking permanent.

Annualizing needs a year to mean anything. Three months of +8% is not +36%/yr. The %/yr curve therefore begins a year into the window, and the toggle is disabled outright on a window that leaves no curve to draw: shorter than a year, or only a year and a day or two, where just one point would remain (the 1Y preset is that case; widen past one year). A year means a calendar year, from a date to the same date a year later, so the 1Y preset's table is annualized and a window a day shorter is not. The table follows the same rule: under a year, CAGR, Sharpe, Sortino and Calmar (and, with a benchmark, excess CAGR and the information ratio) are left blank, while Total and YTD still report. The drawdown track below is unaffected: it keeps measuring real declines from the true window start, including peaks set before the annualized curve begins.

Narrowing the comparison

Sleeve re-runs the backtest on one asset class of each model, renormalized to 100%. This lets you compare two models' bond sleeves head-to-head. Category does the same for one slice category, and combines with Sleeve. A ticker benchmark is not filtered: it stays the whole fund. If a later version of a model holds nothing in the chosen sleeve or category, its curve is held steady from that version until a version holds one again, and the Warnings line says from when. It is never left running on the funds the earlier version held. Explain outperformance decomposes the difference between two models by slice and ticker, over the same window the table measures, so its gap is the difference of the two Total figures. It replays exactly the portfolio each curve does, a slice held steady included. A model built from imported weights has no slices to attribute to, so the button is off for it and says so; rebuild it as a slice-based model to explain it.

Reading the numbers

The Strengths / Trade-offs cards summarize, per model, which metrics it leads and where it trails the leader. This is the same best/worst ranking the table marks, in a sentence. Hover any metric there or in the table for its one-line definition, and see reading the backtest metrics for what each one is for and why two of them can disagree.

The performance and monthly-return tables are named keyboard regions. Tab to the table boundary and use Left/Right to scroll through model or month columns that do not fit at the current window width.

Underlines mark the best (and with 3+ models, worst) value per column; heat coloring is reserved for the risk-adjusted ratios.