Clustering

Also seen as: k-means clustering, regime clustering, k-means regimes

How does a clustering rule decide which days are turbulent?

A clustering rule describes each day by a small set of numbers measured over the observation window, then lets a k-means algorithm find the groups those days naturally form. Each day belongs to the group whose centre it sits closest to:

where $x_t$ describes how much and in which direction the series moved over the window ending on day $t$, $\mu_k$ is the centre of group $k$, $c(t)$ is the group day $t$ is assigned to, and $k_{\text{turbulent}}$ is the group with the highest volatility. In words: the strategy runs Risk-Off on exactly the days that look most like the most volatile stretch of the series' history.

How is clustering different from Hidden regimes (Markov)?

Both rules learn regimes from one series, and they answer differently. Hidden regimes (Markov) gives a probability and lets you set how sure it must be; Clustering gives a yes or no, so there is no threshold to set. Clustering also has no notion of how long a regime tends to last — each day is judged on its own window — so on a series that swings in and out of volatility it can switch more often, and the confirmation delay matters more.

How does Fincanva handle it?

• The series is any instrument you pick under Instrument; a new rule starts on the S&P 500. There is no indicator, operator or pair of thresholds to set, unlike a Single series or Double series condition — see condition types.
• Observation window (trading days) runs from 10 to 63, default 21; the app's note: "How many days to look at to describe each day. 21 ≈ 1 month."
• Number of regimes is 2 or 3, default 2 ("How many groups of days the model looks for."). Whichever you pick, only the most volatile group is Risk-Off; with 3, the middle group stays Risk-On.
• The rule uses only history available on each day. The groups are recalibrated as the backtest moves forward, and a day is never assigned using data from after it.
• It stays Risk-On until it has enough history to learn from, and it stays Risk-On when the groups it finds are too alike in volatility to tell apart.
• One decision, no hysteresis. The confirmation delay is the brake against whipsaw; the app's hint on that field reads "This rule has a single threshold: the delay is your brake against switching too often. 0 = immediate." Auto-rebalance works as it does on any condition.
• Clustering is included from the Advanced plan, at the strategy level and inside a Combined alike; Free and Starter do not offer it. See what each plan includes.
• In the list of risk conditions a saved rule shows "Clustering" with its series, "turbulent group" as its comparison, and its regime count and window, such as "2 regimes" and "21-day window".

What does it look like in practice?

You add a Clustering rule on a broad equity index with the default 21-day window and 2 regimes. Most days of the index's history sit in a calm group — small daily moves, drifting up — and a minority sit in a turbulent group of large moves, mostly down. For months each new day lands in the calm group and the strategy stays Risk-On.

A sell-off starts: over three weeks the 21-day window fills with large down days, and on the day its description moves closer to the turbulent group's centre than to the calm one's, the rule asks for Risk-Off. When the window has rolled past the sell-off and new days land back in the calm group, it asks for Risk-On again. With a 10-day window the same sell-off would register sooner, and brief shocks would register too; with 63 days it would register later and hold longer. The figures are illustrative, not a suggested setting.