KRONOS

A regime-aware quantitative alpha platform — and a market-microstructure research lab — built from first principles.
_{Every model hand-implemented on numpy / pandas / scipy. Every estimator validated against synthetic ground truth before it touches a real price.}

---

KRONOS is two things at once:

1. A production-shaped quant platform — data → regime detection → alpha sleeves → cost-aware portfolio construction → risk overlay → a self-contained interactive dashboard. It runs end-to-end in ~25s and posts a net Sharpe of 1.07 at −17.9% max drawdown — beating the S&P's in-sample Sharpe at just over half its drawdown, sized by a forecast of its own volatility, gated by a fat-tail-aware Student-t regime engine, and tilted by the measured monthly-momentum information budget.

2. A research program that treats markets like physics. 28 pre-registered experiments ask what quant finance genuinely does not know — is volatility rough? how many bits/day does the past leak about the future? are crashes critical transitions or shocks? is the market's near-criticality real? — and answer them with confound-killing methodology, reporting the negative results as loudly as the positive ones.

What ties them together is one discipline: no estimator is trusted until it passes a gate on data where the answer is already known. 38 such gates run in CI. That is the whole point — the platform grades its own homework.

Zero heavyweight dependencies. No scikit-learn, no statsmodels, no PyTorch, no cvxpy. Baum-Welch EM, Student-t HMMs (own ECM), Kalman filters, GJR-GARCH MLE, HAR-RV, fractional-Gaussian-noise simulation, Marchenko-Pastur denoising, Rockafellar-Uryasev CVaR LPs, Hedge learners, deflated Sharpe / CSCV, Hierarchical Risk Parity, Black-Litterman, Hawkes-process MLE, and the canvas charting engine of the dashboard — all hand-built and gate-verified.

Contents

• Quick start
• Highlights
• The platform
• The research program
• Cross-market transfer
• Architecture
• Research integrity
• Reproducibility & data

Quick start

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[data]"          # or: pip install -r requirements.txt

python run_kronos.py              # full pipeline -> output/dashboard.html  (~25s)
python run_research.py all        # 28 research experiments, cached to research/*.json
python run_kronos.py --research   # dashboard with the RESEARCH tab (open output/dashboard.html)
python run_trade.py               # today's research-grounded target portfolio

python tests/run_all.py           # all 38 verification gates (~3min)

Runs fully offline: without yfinance or a network, a seeded synthetic regime-switching market drives the entire pipeline. Force it anywhere with KRONOS_SYNTHETIC=1 (this is what CI uses).

Highlights

• 38 verification gates — 36 proving an estimator has correct size (doesn't fire on null worlds) and power (detects planted effects) on synthetic ground truth, plus 2 calibrating the battery against the real market — all before any real-data claim is made.
• Strictly causal throughout — filtered (not smoothed) HMM probabilities, frozen betas, walk-forward refits, T+1 execution, full transaction costs (commission + spread + square-root impact).
• A rough-volatility replication (H ≈ 0.10) and an information-budget measurement (the daily direction channel is closed; the Sharpe ceiling from sign-prediction alone is 0.48 < beta) — both on the project's own 16 years of data.
• A debunking with teeth: after removing the volatility confound, critical- slowing-down early-warning signals carry no incremental crash-prediction information — the pre-crash signature is ~8× weaker than a real fold bifurcation. Crashes are shocks, not tipping points.
• A cross-market transfer study on Japan / Europe / Asia-EM: the mechanism laws are universal, the frozen US-tuned system holds its risk edge abroad, but the exact law values are local — stated as a clean split verdict.
• A self-audit that found its own sign bug: diagnosing "where does the return go?" exposed an inverted drawdown throttle that 30 green gates had missed (braking at equity highs, releasing into crashes). Fixed under pre-registered kill criteria, shipped with the missing gate, old numbers kept as the baseline row.
• A 1,700-line single-file HTML dashboard with a hand-written canvas charting engine, zero external assets.

The platform

The end-to-end book, run_kronos.py:

prices ─▶ HMM regimes ─▶ regime-gated signals ─▶ HRP + Black-Litterman ─▶ risk overlay ─▶ dashboard
         (Bull/Vol/Bear)  (momentum/rev/low-vol)  (shrunk-cov backbone)   (vol/CVaR/DD)

• Regimes — a Student-t HMM (fat tails modeled within states, per the X² finding; Gaussian engine available as control), refit walk-forward with hysteresis and minimum-dwell stabilization; decisions use strictly filtered (causal) probabilities.
• Signals — 12-1 momentum, short-horizon reversal, and low-volatility, combined with regime-dependent weights (Bull → momentum; stress → low-vol / reversal).
• Construction — a Hierarchical-Risk-Parity backbone on a Ledoit-Wolf / EWMA-shrunk covariance, tilted toward the combined signal via Black-Litterman, long-only with a weight cap.
• Risk — vol targeting as the lever (up to 1.5×, financing charged on the levered portion) with the CVaR cap and drawdown throttle as brakes — the overlay's direction and reach are pinned by their own gate (X27; see KRONOS-EDGE).

Strategy	CAGR	Vol	Sharpe	Max DD	CVaR95
KRONOS (+ momentum tilt, shipped)	+11.8%	11.2%	1.07	−17.9%	1.65%
KRONOS (HAR lever + t-HMM regimes)	+12.0%	11.3%	1.05	−18.8%	1.70%
KRONOS (HAR lever, Gaussian regimes)	+11.7%	11.4%	1.03	−19.4%	1.70%
KRONOS (EDGE baseline: EWMA + Gaussian)	+10.9%	11.6%	0.95	−21.3%	1.76%
SPY (buy & hold)	+15.0%	16.8%	0.91	−33.7%	2.55%

The honest read: KRONOS still does not beat the S&P on raw return — and says so. The edge is risk-adjusted: a better Sharpe at two-thirds of the drawdown, with financing costs on leverage disclosed (1.26%/yr) and the selection-risk caveats (DSR 0.75, PBO 0.45, N=185 logged trials) restated rather than hidden.

The research program

28 experiments, each pre-registered in docs/design/ and gated before real data. The one-line answers — full write-ups, tables, and methods in docs/FINDINGS.md:

Study	The question	The finding
X²	3 regimes or 5?	Gaussian HMMs hallucinate regimes from fat tails — a Student-t HMM stays flat at K=3. It's ~3 regimes + heavy tails.
LAWS	Is there a universal return law?	The One-Clock law: returns are conditionally Gaussian given the realized-vol path (kurtosis 12.6 → 2.6 across 48 assets, one shared distribution).
CLOCK	Is systemic risk contagion?	No — it's correlated volatility clocks. Joint crashes are common vol surges, ~Gaussian copula once you condition on the clock.
SURGE	Does the cascade recurse?	No — it terminates after one level. Volatility has irreducible jumps; the one-clock law does not iterate.
BITS	How predictable is the market?	The direction channel is closed (~0 bits/day, ceiling Sharpe 0.48 < beta); the magnitude channel leaks ~0.4 bits/day.
ARROW	Where does time's arrow live?	In the return↔clock coupling, not in returns themselves — vol-deformation erases the entropy production.
CRITICAL	Critical transitions or shocks?	Shocks. After the vol confound is removed, critical-slowing-down carries no incremental crash signal (~8× weaker than a real bifurcation).
REFLEX	Is the market self-exciting?	Mostly an illusion: 64% of the famous near-criticality (branching 0.68 → 0.25) is volatility clustering, not reflexivity.
CONSTANTS	Do the laws drift over time?	Mechanism constants are constant; only crisis intensity moves (peaks 2020, reverts). No Adaptive-Markets secular drift.
DECATHLON	Smallest market that looks real?	A vol-targeting spiral buys the wild facts; the ceiling is 5/10 and the missing organ is expectation (anticipatory agents).
DECATHLON-2	Is expectation the missing organ?	Refuted. A causal, gate-verified anticipator front-running the vol-targeter flow leaves the ceiling at 5/10 — one layer of expectation moves the sign leak a derivative earlier instead of removing it; information-free prices need the fixed point of anticipation.
TRADE	What system does the science license?	Forecast-vol targeting + regime-gated risk parity + mechanical crash control — risk control, never direction timing.
TRANSFER	Do the laws cross borders?	Mechanism universal, calibration local — see below.
CRYPTO	Do the laws survive outside equities?	Mostly yes — but the leverage effect inverts (crypto +0.03 vs equities −0.04; 8/10 coins flip). Mechanism universal; one law is equity-specific.
FX	Where does FX sit on the leverage triangle?	Statistically zero (+0.005, z vs equities 3.89) — the triangle equity −0.04 / FX ≈ 0 / crypto +0.03 is monotone in microstructure; yen crosses carry the safe-haven tilt.
EDGE	Why is CAGR half the risk budget?	Diagnosis found an inverted drawdown throttle (braking at peaks, releasing into crashes) and an unreachable vol target. Fixed + gated: Sharpe 0.94 → 1.02 unlevered; CAGR 6.4% → 10.9% levered.
EDGE2	What had the research already licensed?	The HAR forecast lever (X28) and the t-HMM engine (X29), each surviving its kill criterion, stacking to Sharpe 1.05 @ −18.8% MaxDD.

Cross-market transfer

Every law above was measured on one universe (48 US tickers). KRONOS-TRANSFER re-runs the entire law battery — and the frozen, US-tuned trading system, with zero re-tuning — on Japan, Europe, and Asia-EM.

A clean split verdict: the mechanism laws (fat tails, leverage effect, near-critical branching, the one-clock collapse) reappear in every market, and the frozen system keeps a positive Sharpe and a shallower drawdown than the local index in all three foreign markets — but the exact law values (H, clock commonality, deformed branching) are market-specific. Universality of mechanism; locality of calibration. The transferable claim is risk control, not alpha.

Frontier: does the leverage law survive crypto?

If the laws are properties of markets, they should survive a market that shares none of equities' plumbing. KRONOS-CRYPTO runs the same 7-law battery on 10 crypto majors — 24/7, no overnight gap, retail-driven, no financial leverage — alongside the equity cohort.

The mechanism is portable — the one-clock collapse (kurtosis 16.6 → 4.5), near-critical branching and its vol-clustering illusion, roughness, and fat tails all reappear. But the leverage effect cleanly inverts: crypto reads +0.03 versus the equity cohort's −0.04 (z = 4.06), and 8 of 10 coins individually flip positive — only BTC and ETH keep the equity sign. The leverage effect is not a market universal; it is a property of the equity microstructure, and it reverses where that microstructure is absent. (Gate X26 licenses the sign reading.)

Architecture

config.py                 every knob; pre-registered parameters
run_kronos.py             v1 pipeline -> dashboard (+ --research tab)
run_research.py           21 KRONOS-X experiments, cached in research/*.json
run_trade.py              the deployable system -> today's target portfolio
kronos/                   37 modules, ~7,500 LOC
  data.py                 prices + OHLC, caching, seeded synthetic fallback
  regime.py / dhmm.py     Gaussian HMM (log-space EM, walk-forward) + duration-HMM
  thmm.py / sjm.py        Student-t HMM (own ECM) + statistical jump model
  signals.py / backtest.py  momentum/reversal/low-vol, regime-gated; T+1 engine
  pairs.py / statarb.py   Kalman pairs + Avellaneda-Lee eigenportfolio stat-arb
  covariance.py / hrp.py  Ledoit-Wolf + EWMA shrinkage; Hierarchical Risk Parity
  black_litterman.py      BL tilt with signal views
  cvar_opt.py / risk.py   min-CVaR LP (Rockafellar-Uryasev); vol/CVaR/DD throttles
  volest.py / vollab.py   Garman-Klass range vol; HAR-RV, GJR-GARCH-t, Diebold-Mariano
  rough.py / rfsv.py      Hurst estimator + fGn simulation; rough-vol forecaster
  laws.py / clock.py      one-clock deformation; correlated-clock tail tests
  surge.py / infobudget.py  cascade structure; KSG/binned mutual information
  entropyprod.py          path-space entropy production (arrow of time)
  critical.py / hawkes.py   critical-slowing-down; Hawkes branching ratio
  constants.py            cross-era law-stability tests
  decathlon.py            agent-based minimal market + the stylized-fact battery
  transfer.py             cross-market law battery + frozen-system backtest
  crypto.py               crypto universe + cross-asset-class leverage contrast
  rmt.py / ensemble.py    Marchenko-Pastur denoising; Hedge/fixed-share learners
  forensics.py            deflated Sharpe, CSCV-PBO, stationary bootstrap
  metrics.py / dashboard.py  performance stats; 1,700-line self-contained HTML
tests/                    38 gates (36 synthetic ground truth + 2 real-data calibration)
docs/                     METHODS, ATLAS, design notes, FINDINGS, research index

Research integrity

The discipline that makes the results worth reading — and a blog-style deep-dive into all of it in docs/METHODS.md ("How do you know you're not fooling yourself?" — the gate philosophy, look-ahead control, the volatility clock, bootstraps, out-of-sample model comparison, overfitting forensics, and information-theoretic ceilings):

• No look-ahead anywhere — filtered probabilities, frozen betas, T+1 execution, walk-forward refits. Look-ahead-sensitive code carries a causality gate (test_trade.py: shifting all inputs forward one day must not change a single historical weight).
• Gate before you claim — every estimator is proven on synthetic worlds with known truth, demonstrating both size (no false positives on null worlds) and power (detects planted effects), before any real-data result.
• Costs everywhere — 1bp commission + 2bp spread + square-root impact (capped), on every trade.
• Negative results reported as prominently as positive ones — stat-arb is dead, RMT doesn't help here, durations don't beat the plain HMM, the trading system doesn't beat the S&P on CAGR. All stated plainly.
• The strategy audits itself — a deflated Sharpe (fed by a trial ledger) and a CSCV probability-of-backtest-overfitting run on KRONOS's own configuration family: the test most backtests never run.

See the Atlas of Ignorance for the open-problem map that scopes the whole program, and CONTRIBUTING.md for the one rule.

Reproducibility & data

Universe: ~48 liquid US equities/ETFs, 2010–2026, Yahoo adjusted OHLC, cached locally (the cache is regenerable and git-ignored). Every random path is seeded; every gate is deterministic. The synthetic market is a seeded regime-switching factor model with Student-t innovations, so the entire platform — pipeline, research, and gates — runs identically offline.

Not investment advice. Backtests on today's liquid names carry survivorship bias and do not predict future returns. This is a research and engineering portfolio, not a trading recommendation.

License

MIT © oliverz