Clawock Glossary

August 30, 2026 · View on GitHub

Source of truth for cross-document terminology. When README, code comments, briefs or PR descriptions first introduce one of these terms, link to its entry here rather than re-translating. EN/ZH parity is enforced at this file by tests/test_glossary_parity.py, not at the translation site.

Each entry has four fields:

  • EN — canonical English term (lowercase; title case only for proper nouns like author names)
  • 中文 — canonical Chinese rendering (one rendering per term; translators do not choose)
  • 一句话 — one-sentence definition a new reader can act on without reading the source
  • First defined — file or doc where the term was first named in this repository

Methodology — factors and signals

EN中文一句话First defined
composite factor复合因子Weighted average of sector-neutral ranks; weights must sum to 1 and equal RAW_FACTORS (adding a component is a re-registration, not a code change).market_data/factors.py::RAW_FACTORS
composite score综合得分The numeric output of the composite factor for one name on one session; the rank the rest of the decision chain sorts by.evaluation/signal_panel.py
composite polarity复合因子极性Diagnostic table showing each component's declared sign vs measured IC sign; classified as polarity_suspect (one component is wrong) or regime (many components are wrong together).PR #1198
polarity suspect极性可疑A single component carries almost all the negative IC — verdict that the factor is broken, not the regime.PR #1198
regime行情状态What this stretch of the market rewards (risk-on / risk-off / trend / mean-revert / high-vol / low-vol), not whether the market is up or down.decision/regime.py
sector-neutral rank行业中性排名Rank within each sector first, then merge across sectors — prevents industry β from contaminating cross-sectional scores.market_data/factors.py
IC (Information Coefficient)信息系数Spearman rank correlation between today's ordering and h-session-ahead returns. ~0.05 industry-floor, 0.10 good, 0.15+ top-tier.evaluation/signal_panel.py
mean IC平均 ICAverage IC across recent sessions; one number, masks more than it reveals.evaluation/signal_panel.py
forward return远期收益Return from today to h sessions later — h ∈ {t1, t5, t20} for the daily bars.evaluation/signal_panel.py
triple barrier三重屏障López de Prado labeling: upper (take-profit), lower (trailing stop), time cap. First touched wins.labeling/triple_barrier.py
chandelier stop吊灯止损Trailing stop at N-day high minus K × ATR; the lower barrier of the triple-barrier scheme.labeling/triple_barrier.py
fixed-horizon固定期Label that closes a position only at h-session close — the old panel column, used as a comparator.evaluation/signal_panel.py
path-aware路径感知Label that exits when the trailing stop is hit intraday — the new panel column.labeling/triple_barrier.py
ambiguous bar模糊 K 线A bar where both upper and lower barriers were touched in the same period; verdict is the unfavorable one.labeling/triple_barrier.py
circular against barrier对屏障循环A signal whose measured IC is mostly restating the barrier rule itself (e.g. stop_distance_pct); flagged CIRCULAR_AGAINST_BARRIER.evaluation/signal_panel.py
quantile structure分位结构Per-session split into three buckets (not five — too few names), then per-bucket forward return averaged across sessions. Splits into top − middle and middle − bottom to identify selection-list vs avoidance-list vs two-sided signals.evaluation/signal_panel.py
persistence持久性Turnover (1 − Jaccard(top bucket) between adjacent sessions) + rank autocorrelation of the signal.evaluation/signal_panel.py
turnover换手率Fraction of the top bucket that differs between two adjacent registration sessions.evaluation/signal_panel.py

Methodology — validation

EN中文一句话First defined
deflated Sharpe (DSR)打折夏普Sharpe minus the expected maximum across n trials, divided by skew-/kurtosis-corrected standard error. Returns insufficient_sample when the underlying sample is too small.evaluation/deflated_sharpe.py
CSCV (Combinatorially Symmetric CV)组合对称交叉验证Split the time series into S groups, run every C(S, S/2) half-half split, observe how often the in-sample winner wins out-of-sample.evaluation/cscv.py
purged CSCVpurged CSCVCSCV with a gap between train and test to prevent information leakage.evaluation/cscv.py
PBO (Probability of Backtest Overfitting)回测过拟合概率The probability that an in-sample winning strategy loses out-of-sample; 0.5 = random, 1.0 = 100% overfit.evaluation/cscv.py
block bootstrap块自举Bootstrap that draws contiguous blocks instead of single rows — necessary when adjacent sessions share a regime. Politis–Romano stationary variant with geometric block length and wrap-around.evaluation/bootstrap.py
BCa intervalBCa 区间Bias-corrected and accelerated confidence interval; uses jackknife to correct both bias and skew. Below MIN_CLUSTERS_FOR_BCA the engine falls back to percentile.evaluation/bootstrap.py
selection rigor选 variant 严格度The combined DSR + PBO + purged CSCV gate; refuses to print a number when any gate says insufficient_sample or insufficient_search.evaluation/add_alpha_walkforward.py
insufficient sample样本不足A diagnostic verdict: the data is too small for the measurement to mean anything; published as a string, not as a number.evaluation/unified.py
insufficient search搜索不足A diagnostic verdict: too few variants were tried for the search to be informative.evaluation/deflated_sharpe.py
grade评级Unified verdict from six gates (cscv / DSR / bootstrap / attribution / drift / …); a refused gate pulls the grade down on purpose. The grade validated is structurally unreachable — that label belongs to rules written before data arrives.evaluation/unified.py
validated已验证Reserved for rules pre-registered before data arrives. Computed grades never return this.evaluation/unified.py
diagnostic诊断性The highest grade a computed evaluation can reach; a working measurement, not a passed gate.evaluation/unified.py
reachability可达性Whether a cohort's two conditions can ever both be true on this book — measured across the whole ledger, not the 30-day window.decision/ledger.py
unreachable cohort不可达 cohortA cohort whose two conditions have empty intersection on the whole ledger — warming_up was the lie, this is the truth.PR #1198

Methodology — liquidity and volatility

EN中文一句话First defined
Amihud illiquidityAmihud 非流动性Absolute return divided by volume as a proxy for the price impact per dollar traded.market_data/bar_signals.py
Roll spreadRoll 点差Implicit bid-ask spread estimated from the covariance of adjacent price changes; valid only when that covariance is negative.market_data/bar_signals.py
Corwin–Schultz spreadCorwin–Schultz 点差Bid-ask spread inferred from intraday high–low ratios; rejects negative estimates from gap days.market_data/bar_signals.py
EWMA volatilityEWMA 波动率Exponentially weighted moving average with RiskMetrics λ = 0.94.market_data/bar_signals.py
GARCH(1,1)GARCH(1,1)Variance-targeting GARCH with refined grid; refuses below 200 observations.market_data/bar_signals.py
realised volatility已实现波动率Standard deviation of recent log returns.market_data/bar_signals.py
realised skew / kurtosis已实现偏度/峰度Higher moments of the return distribution; reported alongside vol-of-vol.market_data/bar_signals.py
vol of vol波动率的波动Variance of recent volatility estimates — a regime indicator, not a return predictor.market_data/bar_signals.py

Methodology — portfolio and risk

EN中文一句话First defined
Ledoit–Wolf shrinkageLedoit-Wolf 收缩Shrink the sample covariance toward a constant-correlation target; the sample-covariance cure for small samples.portfolio/covariance.py
OAS (Oracle Approximating Shrinkage)OAS 收缩A smarter shrinkage target than Ledoit–Wolf; guesses the optimal shrinkage intensity.portfolio/covariance.py
Marchenko–Pastur spectrumMarchenko-Pastur 谱Theoretical distribution of eigenvalues of a pure-noise covariance matrix; eigenvalues outside the boundary are "real signal".portfolio/covariance.py
shrinkage intensity收缩强度How far to pull the sample estimate toward the target; 0 = no shrinkage, 1 = full target.portfolio/covariance.py
effective bets有效持仓数How many independent sources of variance the portfolio actually carries; close to 1 = all in one, close to N = well diversified.portfolio/covariance.py
HRP (Hierarchical Risk Parity)层次风险平价López de Prado 2016: hierarchical clustering + recursive bisection. Used as a yardstick, not as a recommendation.portfolio/allocation.py
Euler risk contributionEuler 风险贡献Each holding's marginal contribution to total portfolio volatility; sums exactly to the total.portfolio/allocation.py
minimum-variance portfolio最小方差组合The portfolio that minimises variance subject to weights summing to 1; here solved by simplex gradient descent because the sample covariance is singular.portfolio/allocation.py
historical stress历史压力Worst observed k-session window with dates attached.portfolio/stress.py
reverse stress反向压力Given a target loss, the most plausible (lowest Mahalanobis distance) scenario that produces it — not "everything falls equally".portfolio/stress.py
Mahalanobis distance马氏距离Distance that accounts for correlations; the reverse-stress search criterion.portfolio/stress.py

Methodology — attribution

EN中文一句话First defined
Fama–MacBeth regressionFama-MacBeth 回归Per-session cross-sectional regression of forward return on factor loadings, then average coefficients across sessions. Requires ≥3 names per factor; refuses to print below that.evaluation/attribution.py
FACTOR_BLOCKS因子块降维Pre-defined grouping of factors into blocks (momentum / stability / quality_liquidity) — a dimensionality reduction, not a search; no other grouping was tried.evaluation/attribution.py
factor return因子收益The mean forward return that a unit exposure to one factor earns across sessions.evaluation/attribution.py
explain_composite复合归因Closed-form decomposition of composite score into per-factor contributions: w_f × rank_f / Σw. Reconstructs the composite to the sixth decimal in tests.evaluation/attribution.py
leave-one-out留一法Removing one factor from every name's composite and re-ranking; answers "would this name's position change if this factor disappeared?".evaluation/attribution.py
tilt / timing倾斜 / 时机Decomposition of common-factor return into a static tilt (allocation) component and a timing (weight-change) component; tilt + timing ≡ common by construction.evaluation/attribution.py
mean R²平均 R²Average cross-sectional R² of the Fama–MacBeth regressions; how much of the return is explained by the factors.evaluation/attribution.py
specific return特异性收益The residual after subtracting common factor returns; in this book, two-thirds of equal-weight returns.evaluation/attribution.py

Decisions and actions

EN中文一句话First defined
decision ledger决策账本The append-only log of every decision the model produced; ground truth for backtests and provenance.decision/ledger.py
tactical entry战术入场An entry decision with a setup and timing; structurally rare — all 7 historical instances pre-date the v1 packet.decision/ledger.py
cut出场信号Decision to exit the position; printed on the scorecard and published in the daily brief.decision/actions.py
hold and watch持有观望Decision to keep the position without sizing up; not the same as "add".decision/actions.py
add加仓Adding to an existing position; rare in this book — see tactical entry.decision/actions.py
shadow portfolio影子组合A parallel book that follows signals without placing orders; used to estimate the cost of execution delay.decision/shadow.py
shadow book影子账本The ledger of the shadow portfolio; lives next to the live book in memory/.decision/shadow.py
mind record决策记录The structured record of one decision — action, condition, regime, size, confidence, evidence.decision/mind_record.py
information packet信息包Bundle of evidence that turns "warming up" into "active"; carries activation_progress for the second gate.decision/packet.py
packet carrying population带 packet 群体The real-world population of decisions that came with an information packet; published separately from the unreachable cohort.PR #1198
warming up预热中Old, dishonest status that pretended a cohort was still filling up; replaced by unreachable_cohort and reachability in PR #1198.PR #1198
activation progress激活进度history_dates tuple on the packet recording how many sessions have passed for each activation gate.PR #1198

Risk and portfolio

EN中文一句话First defined
leverage dial杠杆刻度200-day trend × volatility dial that caps the leveraged-ETF sleeve (×1 / ×0.5 / ×0); a risk-budget control, not a timing signal.decision/regime.py
risk budget风险预算The notional ceiling a sleeve is allowed to risk; published per market per policy.README.md
exposure敞口Dollar or percent-of-book weight of a holding or sleeve; gross or net depending on context.portfolio/risk.py
HHI赫芬达尔指数Concentration measure (Herfindahl–Hirschman Index): Σ wᵢ²; higher = more concentrated; gate renders four bands (green / yellow / orange / red) — only band names are pinned to thresholds; the visual indicator is decided by the dashboard CSS.portfolio/risk.py
covariance协方差Second moment of joint returns; the matrix this engine shrinks via Ledoit–Wolf or OAS before any allocation uses it.portfolio/covariance.py
correlation相关性Covariance rescaled by the product of volatilities; bounded in [-1, +1]; SPCH/SPCX observed at 1.000 in this book.portfolio/covariance.py
conditional number条件数Ratio of largest to smallest eigenvalue; 13,241 on this book — the covariance is nearly singular.portfolio/covariance.py
stress test压力测试A scenario projected onto the book's covariance to estimate loss; here transmits through correlations, not "everything falls equally".portfolio/stress.py

Workspace and book

EN中文一句话First defined
workspace工作区Where the book lives on disk; resolved by workspace_root for any command.workspace.py
engine config引擎配置A config file that ships with the engine rather than with a particular book.workspace.py
book账本One brokerage account + its instrument registry + decision ledger + portfolio snapshot.workspace.py
look-through透传归因Resolving a holding to the issuer whose news, filings and earnings move it (a leveraged ETF to its underlying index).instruments.py

Sessions and calendar

EN中文一句话First defined
session交易日时段One trading day for one market; the atomic unit this engine indexes by.sessions.py
trading day交易日A day on which the market is open for trading.sessions.py
trading session交易时段A resolved trading window (HK afternoon, US pre/regular/post); narrower than session.sessions.py
in session交易中Calendar day and clock both confirm the market is open.sessions.py
closed reason休市原因Short Chinese label (休市, 午休, 节假日) for harness banners.sessions.py
phase session时段映射The calendar session a harness phase belongs to — HK afternoon = pm/close.sessions.py
canonical bar规范 K 线The single resolved bar a session should publish; defined per-market by the canonical raw-bar writer.instruments.py

Reproducibility and provenance

EN中文一句话First defined
run card运行卡A snapshot of one run: seeds, library versions, config files, inputs, wall time, peak RSS, and a metrics_digest.evidence/run_card.py
reproduction key复现键The hash that two runs with identical inputs and code must produce. Library versions are in; interpreter patch and platform are out.seeds.py, evidence/run_card.py
metrics digest度量摘要The hash of the produced metrics; matches the reproduction_key iff the run is fully explained.evidence/run_card.py
unexplained未解释Run state when the key matches but the digest does not — meaning the card still does not describe everything that determined the result.evidence/run_card.py
seed种子A registered random seed; eight modules used to inline literals, now centralised in seeds.py (top-level leaf module — decision cannot import evidence).seeds.py
schema 2schema 2Current run-card schema; deliberately breaks the old key — old keys claimed a coverage they did not have.evidence/run_card.py
wall time / peak RSS运行时间 / 内存峰值Performance dimensions of one run, published alongside numeric metrics so regressions show up in the same evidence.evidence/run_card.py
rows digest行摘要Hash over the fields the scorecard consumes for a set of decisions; binds the scorecard to the rows it claims.scorecard_provenance.py
slice rows窗口行The decisions a windowed scorecard saw, by recorded bounds rather than by clock — for reproducible windowing.scorecard_provenance.py
series digest系列摘要Hash of the whole logical series (archived rows + hot window), not of the working file's bytes — survives cold/hot rollovers.history_store.py

History storage

EN中文一句话First defined
cold / hot split冷热拆分Long-lived rows live in _archive/; recent rows live in the working file. Re-sewn on read.history_store.py
archive path归档路径Where the cold half of a series lives.history_store.py

Architecture and runtime

EN中文一句话First defined
loop循环A phase of the daily pipeline (preflight, run, postflight, watchdog); shared contract lives in _harness_common.py.cli.py
channel通道A delivery surface (Telegram / webhook / etc.); selected by providers/delivery.py.harness/intraday_watchdog.py
runtime运行时The external agent (OpenClaw / Claude Code / Codex / DeepSeek Harness / custom) that owns model calls. clawock is the workflow, not the runtime.cli.py
harnessHarnessThe host process that owns scheduling, retries, gates; lives in harness/.cli.py
command命令A registered clawock <cmd>; the registry lives in utilities.py.utilities.py
utility子命令Same as command; the project uses both names.utilities.py
plugin插件A drop-in feature that adds a command or context surface; lives under plugins/.cli.py
skill技能A packaged prompt + runtime contract; delivered to the model via skills_delivery.cli.py
evidence证据The persisted record (file or run_card) that links a claim to the data behind it.evidence/build_evidence.py
verdict判定A short, decision-facing summary published by the brief or the deck.decision/record.py
context profile上下文画像A pre-validated bundle of capabilities + skills for one kind of run; audited before being used.context/assembly.py
capability root能力根The lazy-discovery root for one capability dimension in an OpenClaw context profile.context/assembly.py
skills delivery技能下发How the model backend receives the skills catalog — full / diff / lazy.context/assembly.py
prompt reportprompt 报告A redacted view of the model-side prompt state for one session, used for parity checks.context/assembly.py
session family会话族The dialect a session key belongs to, ignoring the job or peer id.context/assembly.py

Domain — markets

EN中文一句话First defined
ATR (Average True Range)平均真实波幅Mean of recent intraday true ranges; here ATR(14) sets the chandelier distance.labeling/triple_barrier.py
stop distance pct距止损距离(price − stop) / price; is itself the chandelier distance — circular against barrier when used as a signal.labeling/triple_barrier.py
look-through exposure透传敞口Gross and HHI exposure computed on the underlying issuer of a holding, not on the wrapper.instruments.py
benchmark基准The market index used for residualisation and regime comparisons.market_data/benchmarks.py

Conventions used here

  • One EN term → one 中文 rendering. Translators do not pick alternates.
  • First defined points to the earliest file or doc that defines or uses the term; if later code is the authoritative source, this column still points to the originating module to keep history honest.
  • "Reject" / "refuse" / "print" are used deliberately: many of these measurements have a refusal path that returns a string verdict instead of a number, and that is part of the contract.
  • Code words and identifiers stay in their original case; bold is reserved for emphasis inside one-line definitions, not for term names.