Market regime analysis
Segmenting a stretch of market history into distinct environments (“regimes”) — each with its own volatility, correlation, and leadership character — and adapting to which one you’re in rather than to a single running trend. Smolinski calls it the most interesting part of his after-action-review: he split H1 2026 and early July into five regimes — a low-vol grind, a ~9% geopolitical drawdown, a V-shaped recovery, a June mega-cap-led pullback, and a dispersed lower-correlation rebound smolinski-h1-2026-aar.
The point
The signal wasn’t any one regime but the speed of transitions — four-to-five very different environments in a compressed window. The operational lesson: stay calm when volatility rises, and distrust decisions anchored to longer-term trends that are themselves in flux. Two regime readings from the same deck:
- Leadership breadth — the “Mag 7 drives everything” regime went out of vogue; dispersion widened within large/mega-caps and across market-cap segments, with small caps leading.
- A regime-flip risk — a renewed rate rise (futures priced a possible September Fed hike) could curb small-cap momentum and restore mega-cap/tech leadership, i.e. flip the breadth regime back.
The automated counterpart (2026-07-26)
Koshtenko’s robot derives regimes the other way — Gaussian Mixture Model clustering over feature vectors, feeding the cluster as context to an XGBoost signal model — and he wants to go further: ~100 regime-specialized models instead of one ensemble. So the spoke now holds both readings of the same concept: a human segmenting history into five named environments from a chart, and an unsupervised model partitioning feature space into clusters with no names at all. The hand-made version is interpretable and unfalsifiable; the clustered version is reproducible and uninterpretable. Neither source tests whether its regimes predict anything out of sample.
Relation to the automated stack
Regime analysis is a market-state concept, upstream of the predict/decide/execute layers this spoke’s systems automate: a regime label is context a strategy conditions on. Smolinski derives his by hand from freely available data (yfinance prices, FRED rates, the Cboe SKEW tail-risk index); the automated analogue is a model that infers regime as a latent feature. The spoke has no such source yet — this is the first regime-aware page and a marker for where a systematic regime-detection source would slot in.
A regime shift is also what defeats a correct backtest. chan-algorithmic-trading names historical episodes where “even the most correct backtest will fail to predict the future returns of a strategy”, and gives a concrete mechanism from the mean-reversion side: a profitable ETF pair unhinged by a fundamental change, where the statistical relationship was real and then stopped being true. That connects this page to backtesting directly — regime shift is the one item on Chan’s pitfall list that no amount of implementation care can remove.
A regime gate as a shipped component (added 2026-08-03)
This page has treated regime as a diagnosis — the thing that breaks a working strategy.
superior-skills ships it as a reusable component: regime-overlay, a triple-confirmation
gate combining EMA separation, ADX and an N-bar return, sold as what “turns fragile directional
strategies into regime-robust ones.” Its risk disclosure goes further and makes gating mandatory
advice — “strategies without regime confirmation are demonstrably fragile.”
It also states a structural claim this page can use: two strategies chosen to be regime-complementary and run as separate sub-accounts, where the trend-following gate fires zero trades during the chop windows in which the mean-reverting strategy works, and underperforms during the trends the first captures. That is regime analysis used for portfolio construction rather than for a post-mortem — pick strategies whose live windows don’t overlap, and the gate becomes the allocator.
Unverified, and from a vendor: the claim rests on the same single 162-day window as the returns, and one of the two strategies has six trades in it (backtesting). The idea is testable and worth more than the numbers attached to it.
Related
chan-algorithmic-trading · mean-reversion · backtesting · smolinski-h1-2026-aar · after-action-review · erik-smolinski · algorithmic-trading · superior-skills · agent-deployed-trading