Stress test

Also seen as: stress testing, scenario analysis, what-if analysis, entropy pooling

How can a stress test work without inventing data?

By changing how much each past day counts rather than what happened on it. Normally every day of the backtest counts equally. To make the history average −20% a year, the test counts the losing days a little more and the winning days a little less — just enough for the weighted average to hit the target, and no more. To double volatility, it leans on the large days, up and down, over the quiet ones. The textbook name is entropy pooling: among all the ways of reweighting the days that satisfy the scenario, it picks the one closest to the original, equal weighting. Every figure it reports is still made of days the strategy really lived.

That is also its limit. A scenario can only be reached with the kinds of day the history contains: a strategy that never had a bad day cannot be made to average −40% a year, and the panel then says "We couldn't calculate this scenario."

How do I read the two gauges?

They are the honesty check on the result, and they are the first thing to read.

• How far we bent history measures how different the reweighted history is from the original. "A little" means the scenario is plausible for this strategy; "Noticeably" means it asked visible effort of the history — read the result with care; "A lot" means the scenario is far from anything the strategy has lived through, and the result is not very reliable.
• Days that still count measures how much of the history still carries real weight. It reads, from best to worst: the result rests on almost the whole history; on a good part of it; on a small part of it — read it with care; on a few extreme days — read it with great care.

A result with both gauges in their first reading is the one to lean on. When the worst 5% of days shrinks to a single day, the VaR and CVaR rows are marked "· a single day": the two are then equal and say little.

How does Fincanva handle it?

• It runs on request, from the Robustness tab's Stress test panel. Pick one of the Ready-made scenarios — "A −20% year", "Double volatility", "A flat year", "Crash: −40% and volatility ×2" — or set Assumed annual return (−50% to +50%) and Volatility multiplier (0.5× to 3×) yourself, then press Apply scenario. At least one of the two has to differ from normal; a volatility multiplier alone keeps the average return where it was.
• The result is a table, On the strategy's historical days, with the Base figures beside With the scenario: average annual return (arithmetic), annual volatility, daily VaR 95% and daily CVaR 95%. VaR and CVaR are daily losses, shown as positive numbers; they are defined on Performance Metrics.
• The Base column is computed on the same footing as the scenario column, not the way Performance Metrics computes its figures, so it does not match that page. Compare Base with With the scenario.
• The figures follow the tab's Simulation settings: with Costs & interests or Taxes on, the test runs on the net returns.
• A simulation stored before the panel existed has to be run again first.

The stress test is included from Advanced upwards, on the same plans as backtest reliability, so the Robustness tab opens as one. See what each plan includes.

What does it look like in practice?

A strategy's history averages 8% a year with 15% volatility, and a worst-5% day loses 1.4%. Apply "A −20% year": the table shows the average return moved to −20%, volatility up slightly to 17%, and the daily VaR at 2.1% — the reweighted history leans on its bad days, and the bad days are bigger than the average ones. How far we bent history reads "Noticeably" and Days that still count says the result rests on a good part of the history: a plausible, if demanding, scenario for this strategy. Try "Crash: −40% and volatility ×2" on the same strategy and both gauges drop to their last readings — the history holds too few days like that for the answer to mean much.