The strategy catalog — twenty-nine panel/event shapes, one ranking helper

alphakernel registers twenty-nine strategies as of writing. Twenty-eight of them share a single 17-line cross-sectional ranking helper (_xs_weights_from_scores, ADR-0057) and differ only in the score function or the per-tick gate. This note is the catalog — what’s shipped, how the shapes relate, and where to point a runner.

The runner reads slug from a strategy’s class attributes; everything else (cites, feature_slugs, intent) lives in the provenance row. ak strategies list surfaces the full catalog with one command.

Panel-clock signals

These read prices on the bar cadence; no event awareness.

Slug	Score	Note
equal_weight_long	1/N per priced symbol	Pipeline exerciser. intent="exerciser"; no thesis.
xs_momentum_live	log(P_t / P_{t-lookback})	The unit — single-factor momentum, top/bottom quantile sorted.
xs_long_only_momentum	top-quantile only	Long-only variant for operators who can’t short (cash accounts, IRAs, mutual-fund mandates). Default gross_leverage=1.0.
xs_vol_weighted_momentum	log(P_t / P_{t-lookback}) / std(returns)	Vol-scaled momentum — risk-adjusted version of the unit. Often the top arm in real-data tests.
xs_momentum_with_spread_filter	XsMomentum + chop filter	Only fires when top-bottom spread > threshold. Closes the chop hole the strategy-shape-beats-factor-count note named.
xs_momentum_with_vol_regime_filter	XsMomentum + per-symbol vol-regime gate	Drops symbols whose vol_5/vol_60 exceeds threshold. Scale-invariant gate per vol-ratio-as-cross-symbol-regime-gate. Tops mean Sharpe on the harness; pays in variance per vol-regime-filter-mean-vs-variance.
time_series_momentum	per-symbol `	log return
xs_three_clock_momentum	linear combination of mom_5, mom_20, mom_60	Composite-horizon momentum. Default weights (-0.5, +1.0, +0.5) express short-reversion + medium-trend + long-confirm. Highest min-Sharpe across seeds; see three-clock-momentum-tops-the-harness.
xs_three_clock_momentum_with_vol_regime	three-clock score + vol-regime gate	Composite of the two top arms. INTERFERES on single-signal data, STACKS on dual-signal — see composite-strategies-can-interfere and dual-signal-makes-composites-stack.
xs_mean_revert	-log(P_t / P_{t-lookback}) over short lookback	Sign-flipped momentum. Short-clock cross-sectional reversion. Diagnostic arm for the dual-signal harness mode (PR #681) — picks up short-clock mean-reversion when present, penalised when not.
xs_momentum_with_vol_regime_transition_filter	XsMomentum + per-symbol vol-CHANGE gate	Drops symbols whose |vol_5_now / vol_5_prev - 1| exceeds threshold. Fix attempt for vol_regime_filter’s clustered-vol underperformance; disconfirmed (see disconfirmed-the-transition-gate-fix).
xs_momentum_with_vol_penalty	XsMomentum − vol_penalty × vol_5/vol_60	Score-stage continuous vol penalty. Interpretation B from PR #723 — different intervention point than the filter variants.
xs_momentum_with_portfolio_vol_gate	XsMomentum × portfolio_scale(median vol_5/vol_60)	CROSS-SYMBOL vol intervention: universe-wide median vol sizes the whole portfolio down. Recovers 75-80% of the Sharpe gap per-symbol variants lose (PR #737).
xs_three_clock_momentum_with_portfolio_vol_gate	three-clock × portfolio_vol_gate	Composite at MIXED stages (per-symbol × cross-symbol). Interferes at +1.00 correlation with three_clock (gate dormant on this data); see perfect-correlation-explains-the-interference.
xs_three_clock_vol_weighted_momentum	three-clock score / per-symbol std	Composite at SAME stage (score × score) of the most-decorrelated pair (+0.71). At N=10 beat both parents; at N=100 ties baseline (see hundred-seeds-confirms-and-converges).
xs_time_series_filtered_momentum	ts-momentum filter + cross-sectional rank	Composite:
xs_time_series_active_set_momentum	ts-momentum filter + sign-preserving magnitude weighting	Composite: same filter, weights = sign × relative magnitude. Recovers the sign discipline ts_filtered broke. Recovers most of ts_momentum’s edge without stacking above it.
xs_ranked_momentum_with_vol_threshold	XsMomentumLive + per-symbol absolute-vol floor	Pure rank-based + absolute-vol filter (no sign conflict). Pairwise-rule prediction: behaves like baseline. Empirically confirmed at N=100: +0.830 vs baseline +0.863, within stdev.
xs_momentum_ma_crossover	(MA_short - MA_long) / MA_long	Different score-stage shape from raw return — smooths over per-day noise. N=100: +0.742 (below baseline +0.863) with lower stdev (1.033 vs 1.077) and better worst-case.
xs_vol_weighted_mean_revert_composite	momentum_weight × (mom/std) + revert_weight × (-short_clock_return)	Score-stage composite of anti-correlated parents (-0.42). Predicted to stack; empirically ties vol_weighted parent (anti-correlation hedges, doesn’t stack — see anti-correlation-hedges-not-stacks).
xs_momentum_live_with_ma_composite	mom_weight × log(P/P_lookback) + ma_weight × (MA_short - MA_long)/MA_long	Score-stage composite at +0.70 correlation (ADR-0062 marginal band). First robust single-signal stack: +0.071 above baseline at N=100 (see the-first-robust-single-signal-stack). Redundant-measurement composition: both parents see 20-day momentum at different smoothing scales.
xs_vol_weighted_three_clock_composite	vw_weight × (mom/std) + tc_weight × (multi-horizon momentum)	Same +0.70 correlation as mom_ma_composite but parents measure DIFFERENT quantities (vol-adjustment vs horizon-stacking). INTERFERES below both parents (multiplicative interference per ADR-0062’s 5th condition; see redundant-vs-multiplicative-composition). The empirical contrast that confirmed the 6th iteration.
xs_two_factor	mom + w_z × zscore	Adds mean-reversion. Default w_z = -1 makes it reversion-dominant.
xs_three_factor	mom + w_z × zscore + w_s × skew	Adds tail-skew aversion. Default w_s = -1 shorts positively-skewed names.
xs_four_factor	mom + w_z × zscore + w_s × skew + w_k × kurtosis	Adds tail-fatness aversion. Default w_k = -0.5; see higher-moments-add-noise-faster-than-signal for tuning.

Each non-exerciser strategy cites walk-forward-without-leakage for the live-cadence discipline and the-deploy-contract-isnt-a-yaml-file for the citation interface. All four use the same ranking helper — the score is the only thing that changes.

Event-aware wrappers (around the panel signal)

These read panel prices but consult an event_release silver table to gate or modulate exposure around scheduled releases.

Slug	Shape	Window	Mode
xs_momentum_with_fomc_blackout	Binary gate, flat-inside	24h pre-event	Risk reduction
xs_momentum_post_fomc_drift	Binary gate, trade-inside	4h post-event	Alpha capture (continuation)
xs_momentum_post_event_reversal	Binary gate, trade-inside (sign-flipped)	4h post-event	Alpha capture (mean-reversion)
xs_momentum_with_event_damping	Continuous gate	Exponential 1 − e^(−τ/half_life) around any release	Risk reduction (smooth)

All four wrappers accept release_kind_filter (default "fomc", or None for any-kind, or "earnings" etc.) so a new event-kind adapter automatically extends the wrapper coverage without class changes.

The drift and reversal wrappers are direct counterfactuals — see drift-vs-reversal-as-counterfactual for the discipline rule of running both arms to measure whether the post-event regime is continuation- or reversion-flavored.

All three cite the-event-clock-isnt-the-panel-clock for the surface and one or more of the discipline notes (dont-pay-for-caution-you-cant-justify, the-baseline-arm-you-forgot, event-aware-needs-signal-concentration) for the measurement bar they have to clear. The damping strategy is release-agnostic; the blackout and drift are FOMC-only by class (release-agnostic versions are a parameter away — the existing filter_fn shape supports it).

Per-tick citation: each wrapper sets cited_event_artifacts = ("event:event_release/<id>",) while inside its gating zone, cleared outside. The ADR-0031 cooldown rule reads that citation when deciding whether to gate.

The decomposition

Two orthogonal axes — score shape (how do we rank within the universe?) and entry rule (when do we trade at all?). Most of the catalog varies on the score axis; time_series_momentum and xs_momentum_with_spread_filter are the catalog’s two entry-rule shapes.

Score axis

	Panel-only score	Event-aware gate
Long-only single-factor	xs_long_only_momentum	(compose with any wrapper)
Single-factor	xs_momentum_live, xs_vol_weighted_momentum	…_with_fomc_blackout, …_post_fomc_drift, …_post_event_reversal, …_with_event_damping
Multi-window single-factor	xs_three_clock_momentum	(compose with any wrapper)
Mean-reversion (short-clock)	xs_mean_revert	(compose with any wrapper)
Composite (score + gate)	xs_three_clock_momentum_with_vol_regime, xs_three_clock_momentum_with_portfolio_vol_gate	(compose with any wrapper)
Composite (score × score)	xs_three_clock_vol_weighted_momentum, xs_time_series_active_set_momentum	Sign-preserving same-stage composition
Composite (filter × rank)	xs_time_series_filtered_momentum	Sign-vs-rank conflict; documented failure mode
Two-factor	xs_two_factor	(compose with any wrapper)
Three-factor	xs_three_factor	(compose with any wrapper)
Four-factor	xs_four_factor	(compose with any wrapper)

Entry-rule axis

Entry rule	Strategy	Notes
Always trade quantiles	every Xs* score variant above	Default cross-sectional shape.
Trade only when top-bottom spread > threshold	xs_momentum_with_spread_filter	Closes cross-sectional chop.
Trade only when per-symbol vol_5/vol_60 under threshold	xs_momentum_with_vol_regime_filter	Per-symbol vol level gate. The original “clustered-vol breaks the gate” claim (vol-cluster-breaks-the-regime-filter) was small-N noise — see the-clustered-vol-finding-was-also-small-n for the N=30 re-run that disconfirms it. The gate still loses ~0.21 Sharpe to baseline on i.i.d.-vol data at N=100.
Trade only when per-symbol |Δvol| under threshold	xs_momentum_with_vol_regime_transition_filter	Per-symbol vol change gate. Disconfirmed fix (PR #719).
Size portfolio down when universe-wide vol elevated	xs_momentum_with_portfolio_vol_gate	CROSS-SYMBOL intervention. Best vol-aware variant.
Score-stage continuous vol penalty	xs_momentum_with_vol_penalty	Per-symbol continuous; doesn’t drop symbols.
Trade per-symbol only when |absolute return| over threshold	time_series_momentum	Closes time-series chop.

The wrappers compose orthogonally with the multi-factor scores: XsMomentumWithFomcBlackoutStrategy(inner=XsTwoFactor(...)) is a trivial follow-up if the operator wants a two-factor signal plus a pre-event flatten. The shipped wrappers only wrap XsMomentumLive because that’s all that’s been needed so far — the seam is there to widen when the question warrants it.

How an operator picks

The designing-an-event-aware-strategy checklist names the six steps. The short version:

1. Panel-only signal works on the data? If no, fix the signal first — wrappers can’t add alpha to noise.
2. Is there an event that should change exposure? If yes, the wrappers are the next move. If no, two-factor or three-factor may.
3. A/B all of it. The 31-arm harness in examples/fomc_blackout_compare.py runs every strategy in this note against the same synthetic data in one command. Add real data when it lands.

For the cross-arm findings the harness has surfaced so far — including the stack-vs-interfere result for composites and the mean-vs-variance trade for the regime gate — see what-24-arms-told-us. It’s the session-summary index over the per-strategy notes.

Synthetic-equivalence note (real data may differ)

At N=100 seeds on the default single-signal synthetic, several catalog entries are statistically indistinguishable (see hundred-seeds-confirms-and-converges):

Equivalence class	Arms	Reason
baseline ≈ xs_momentum_with_spread_filter	both at +0.863	chop filter rarely fires on this synthetic
xs_three_clock_momentum ≈ xs_three_clock_momentum_with_portfolio_vol_gate ≈ xs_three_clock_vol_weighted_momentum	all at +0.798	universe vol rarely crosses threshold; vol-weighting marginal

The arms are NOT identical — they intervene at different stages (score, portfolio scale, score-normalisation). On REAL data with active vol regimes and cross-symbol correlation structure, the variants should diverge. This catalog keeps all of them because:

1. Synthetic-equivalence is data-specific. A future alphakernel run on Polygon bars might find the chop filter firing 15% of the time (where the synthetic has it firing 0%), separating spread_filter from baseline.

2. Deprecating a shipped strategy sends a stale-thesis signal per ADR-0014. The strategy doesn’t claim to work; the catalog claims to have it available.

3. Operators picking among equivalents can read the shape (filter vs portfolio scale vs score-normalisation) and pick the one whose mechanism they trust on their real data.

What this catalog is not

• Not a recommendation set. Every strategy here ships with citation discipline (ADR-0011/0015) but the platform makes no claim that any of them currently captures alpha. The A/B harness is the place to ask that question; the catalog is the working surface.
• Not exhaustive. A risk-parity cross-sectional shape (XsRiskParityStrategy), a pairs-trade shape, a regime- switching wrapper — natural follow-ups, none shipped. The catalog grows as the research-question pipeline produces specific shapes worth registering.
• Not a leaderboard. Operators sort the catalog by their own question; the platform sorts by registration order. The ak strategies list command is alphabetical by class name.

What gets added next

Whatever the next research question demands. The platform’s strategy registration cost is now ~150 LOC + tests + an entry in __all__ + a row in the citation graph. The bar for “is this question worth registering a strategy for?” is the same bar the dont-pay-for-caution-you-cant-justify note names — does the question have a measurable answer, and does the answer warrant a separate shipped artifact?

If yes, the strategy ships. The catalog grows. The 6-arm harness gets another arm. The citation graph stays honest.