Min CVaR is an allocation method that chooses the weights whose average loss on the worst days is as small as possible. It looks only at the bad tail, the worst 5% of days by default, so it separates instruments variance calls equally risky when one has rare, deep falls. Scenario CVaR applies it to generated futures.
Also seen as: minimum CVaR, mean-CVaR optimisation, minimum expected shortfall, CVaR minimisation
How does Fincanva handle it?
The method picker labels it "Min CVaR" and describes it as "Reduces the average loss on the worst days"; Scenario CVaR as "Generates possible futures and reduces the worst losses across them".
- Min CVaR and Scenario CVaR are offered at both levels: across the instruments of a strategy, and across the strategies of a Combined.
- Both read the calculation window (In-sample, 12 months by default). Neither reads the Covariance matrix choice: they work on the returns themselves.
- The Tail share starts at 5% and accepts 1% to 25%; a value outside that range shows an error on the field, and the backtest does not run until it is corrected.
- Both are marked "slow to compute" in the picker: each rebalance solves an optimisation over every day or scenario.
Which plan includes Min CVaR and Scenario CVaR?
Both depend on your plan, at each level where they are offered. The robustness radius and joint-crash scenarios come with Scenario CVaR itself.
Min CVaR inside a strategy
Included from Ultimate upwards. See what each plan includes.
Min CVaR inside a Combined
Included from Ultimate upwards. See what each plan includes.
Scenario CVaR inside a strategy
Included from Ultimate upwards. See what each plan includes.
Scenario CVaR inside a Combined
Included from Ultimate upwards. See what each plan includes.
What does Min CVaR minimise?
The conditional value at risk (CVaR, also called expected shortfall): the average loss across the worst share of outcomes, that share being the Tail share.
- the average loss on the worst α of days — what the method pushes down
- the portfolio's loss on one day
- the loss that only the worst α of days reach, α being the tail share
Line up every day of the window from worst to best, keep the worst 5%, and average them. Variance counts a +3% day and a −3% day as the same risk; CVaR only counts losses, which is why two instruments with the same volatility can come out very differently.
What is the tail share?
The Tail share is Min CVaR's one setting: "The worst share of days or scenarios over which the method measures the average loss." It runs from 1% to 25% and starts at 5%. A small share looks at the few most extreme days only, which makes the estimate noisy; a larger share is steadier but looks less deep into the tail.
What is Scenario CVaR?
Scenario CVaR minimises the same average loss over generated possible futures instead of the window's days. At each rebalance it builds a set of scenarios, each a possible next period assembled from real days of the window, and minimises the average loss of their worst share. Two settings describe them, under Advanced settings:
- Number of scenarios — "How many possible futures are generated at each rebalance." From 100 to 2,000; 500 by default.
- Horizon of each scenario — "How many trading days each generated future covers." From 5 to 63 days; 21 by default, about one month.
How does Scenario CVaR differ from Min CVaR?
In two ways. The loss it looks at is over a whole next period rather than a single day, and the scenarios can combine days that never followed each other in history — two instruments that never fell on the same day can still fall in the same scenario. The limit is the other side of the same coin: a kind of market the window never contained cannot appear in any scenario.
What does the robustness radius do?
The Robustness radius makes Scenario CVaR also guard against scenario sets slightly different from the one it generated: "Also protects against scenarios slightly different from the generated ones. The higher it is, the more evenly the weights spread." Off by default; switched on, it runs from 0.001 to 0.1 and starts at 0.01. In theory terms it is distributionally robust optimisation: an adversary may shift the scenarios by the radius before the loss is measured, and the method minimises the worst loss that could result. Concentration is penalised, so from a radius of about 0.02 the weights become almost equal and every other setting stops mattering; with joint-crash scenarios also on, the app says so: "At this radius the weights become almost equal", with an action to reduce the radius to 0.01.
What are joint-crash scenarios?
Joint-crash scenarios generate the futures differently: each instrument keeps its own history of daily moves while the link between instruments is drawn so they crash together more often than the window shows — a Student-t copula. The app puts the trade-off plainly: "t copula: instruments tend to crash together. It is a different way to generate scenarios, not necessarily a better one." It gives up the way calm and stormy days cluster in real history. Switched on, it adds Tail heaviness, from 3 to 30 and 5 by default: "Lower = more extreme joint crashes."
What does it look like in practice?
Two instruments both have 15% annual volatility over a 12-month window of about 252 days. A moves about ±0.95% on a typical day; B is calmer, about ±0.57%, but had four days of −6%, which supply almost two-thirds of its variance.
- At a 5% tail share the method averages the worst 252 × 5% ≈ 12–13 days.
- A's worst days average about −1.95%; B's — its four −6% days plus eight or nine ordinary days of about −1.25% — about −2.8%.
- Variance sees two equally risky instruments; Min CVaR sees about 40% more tail loss in B and leans toward A, holding some B only where B's bad days fall on A's good ones.
What happens when the window is too short for the tail share?
When the in-sample window times the tail share is less than one day, the average becomes the single worst day. The app warns before you run — "Too little data for this tail share" — with the arithmetic: at 1 month and 4%, about 21 days × 4% is less than one day (at 5%, 1.05 days, right at the edge). It offers two fixes: lengthen the in-sample period, or "Switch to EVaR" — Entropic value at risk, to which every day contributes. Scenario CVaR measures its tail over scenarios, so the check runs on their number instead, and within the accepted ranges its tail always keeps at least one scenario.