SKILL.md
Regime — Markov regime detection for any asset
This skill answers one question for any asset: what regime are we in, how sticky is it, and what does that imply for risk and direction? It's built to slot into a trading agent the user already has — as a confirmation layer, a signal, or a risk gate — without them rewriting their strategy.
Framework: Roan (@RohOnChain). Refactored into this plugin by Lewis Jackson. Backtests are historical, not forward-looking.
How to invoke
One command. It takes EITHER a ticker OR the user's own CSV, so it drops into any pipeline regardless of asset:
# any ticker yfinance knows (stocks, ETFs, crypto, FX, futures):
uv run ${CLAUDE_PLUGIN_ROOT}/scripts/markov_regime.py --ticker BTC-USD --json
# the user's own price file (their data, their asset, their pipeline):
uv run ${CLAUDE_PLUGIN_ROOT}/scripts/markov_regime.py --csv ./my_prices.csv --json- Drop
--jsonfor the on-camera pretty terminal output (matrix, persistence
diagonal, stationary mix, walk-forward Sharpe + max DD, HMM line).
--csvneeds only a date column and a close column. It auto-detects common
names (date/time/timestamp, close/adj close/price/last); if there's exactly one numeric column it uses that. No reformatting required.
- Defaults:
--window 20,--threshold 0.05(±5%),--years 10,
--min-train 252. All overridable. --no-hmm skips the HMM.
- First
uv runresolves dependencies once (~10–20s), then it's instant.
The JSON contract (every field)
--json prints exactly one JSON object to stdout and nothing else. On failure it prints {"error": "..."} and exits non-zero. Fields:
| Field | Type | Meaning | |
|---|---|---|---|
source |
str | the ticker or CSV path analysed | |
rows |
int | number of price rows used | |
datestart, dateend |
str | ISO dates of the window | |
params |
obj | window, threshold, min_train actually used |
|
states |
list | ["Bear","Sideways","Bull"] — fixed index order 0,1,2 |
|
current_regime |
str | the regime as of the last bar | |
nextstateprobabilities |
obj | bear/sideways/bull — P(next \ |
current) |
signal |
float | bullprob − bearprob in [-1, 1]. >0 long bias, <0 short bias, magnitude = conviction | |
transition_matrix |
3×3 list | row = from-state, col = to-state, rows sum to 1 | |
persistence_diagonal |
obj | bear/sideways/bull — P(stay in same regime). High = sticky regime |
|
stationary_distribution |
obj | bear/sideways/bull — long-run fraction of time in each regime, sums to 1 |
|
walk_forward |
obj | sharpe, maxdrawdown, ntrades from a re-estimated-every-step, no-lookahead backtest. sharpe/max_drawdown may be NaN if history is too short |
|
hmm |
obj | available: true → regimes (label, latentstate, meandaily_return) + caveat. available: false → reason (graceful degrade — everything else is still valid) |
|
framework, disclaimer |
str | attribution + "historical, not forward-looking" |
signal and currentregime are the two fields most strategies consume. stationarydistribution is the one most risk layers consume.
Composition — slot it into what the user already has
The user already has a trading agent or strategy on some asset. This skill is a layer they add, not a system they adopt. Run it, read one or two fields, gate their existing logic. Three patterns:
(a) Regime confirmation on an existing momentum/strategy
The user has entry logic that already fires. Wrap it: only take longs when the regime agrees, only short when it disagrees.
import json, subprocess
r = json.loads(subprocess.check_output(
["uv","run",f"{PLUGIN}/scripts/markov_regime.py","--ticker","SPY","--json"]))
if my_strategy_says_long and r["signal"] > 0:
enter_long() # momentum + regime agree → take it
elif my_strategy_says_long and r["signal"] <= 0:
skip() # momentum says go, regime says don't → stand downOne line of gating. Their strategy is untouched; the regime just vetoes trades that fight the prevailing chain.
(b) Stationary distribution as a tail-risk / position-size filter
The stationary mix is the asset's long-run baseline. A high baseline Bear share means this asset structurally spends a lot of time in drawdown — size down.
bear_baseline = r["stationary_distribution"]["bear"]
size = base_size * (1.0 - bear_baseline) # heavier bear regime → smaller bets
# or hard gate: if bear_baseline > 0.40: size = 0 # too tail-heavy to tradeNo new model. The user keeps their sizing logic and scales it by a single number that reflects how regime-dangerous the asset actually is.
(c) Standalone signal
No existing strategy needed. The signal field is already a direction + conviction in [-1, 1]:
position = r["signal"] # +0.6 → 60% long; -0.4 → 40% short; ~0 → flatSanity-check it first with the printed walk-forward Sharpe + max drawdown (run without --json to see them on screen) before sizing real capital.
Notes
- Asset-agnostic by design.
--tickerfor anything yfinance covers;
--csv for the user's own data on any asset/timeframe their pipeline produces. The math is identical either way.
- No lookahead. The walk-forward refits the matrix using only past data at
every step (incremental O(n), bit-identical to a from-scratch rebuild).
- HMM degrades gracefully. If
hmmlearncan't compile (e.g. Windows
without MSVC), hmm.available is false with a reason and every other field is still correct. HMM states are labelled by ascending mean return, so a positive "Bear" mean just means the worst latent state was still net-positive over that window.
- Defaults match the on-camera demo and the Pine Script bonus (window 20,
±5%). Use --threshold 0.02 to reproduce the tighter labelling from the original onboarding prompt.