daily-seasonal

Imported from mslade50/New_Seasonals (.claude/skills/daily-seasonal/SKILL.md). Install upstream with npx skills add mslade50/New_Seasonals --skill daily-seasonal. Copyright stays with the author.

Daily Seasonal

Deliver up to three seasonal trade ideas, found on the whole seasonal surface and interrogated against the data before they reach McKinley. He reads them, says yes or no per idea, and places orders. There is no conversation: the ideas must be finished when they arrive.

The run is 07:00 ET on weekdays, after the Daily Pitch has published. Three is the full slate. One or two ship when that is what survived, with the empty slots paid for; a morning where nothing survives ships a stand-down. Never pad, never publish nothing.

Design of record: docs/seasonal_agent_design_2026-09-30.md. Where this file is silent, the Daily Pitch skill (.claude/skills/daily-pitch/SKILL.md) governs.

What this is

A falsification desk that owns the seasonal surface end to end. The board ranks cells; this agent tries to kill trades. On 2026-09-30 all four board tickets died under the pitch battery: WMT and GS were sector beta, TXN flipped sign at the tradeable entry, and TRV was a real 26-year pattern, decayed since 2016, sitting in its losing regime branch.

What this is not

• Not the site board. The board is a scanner and, for this agent, a candidate feed. It stays live on the site. A board ticket is a candidate, never evidence: every one is re-checked from scratch from the entry the order would actually fill at.
• Not the systematic book. A trade that is materially a St OS Sznl or Weak Close Decent Sznls signal (see book.staged_signals) is dead on arrival unless it adds a real twist (instrument, legs, side, or a meaningfully different window), and the write-up names the overlap.
• Not the pitch. An idea matching history.pitch_recent_fingerprints (what the pitch shipped inside 10 sessions) is dropped, or carries a changed_since sentence saying what materially changed. The pitch negative registry is read-only here; this product keeps its own.
• Not a place orders happen. Nothing is placed without McKinley's Y on the tab.

False positives are fine. He filters. Unchecked ideas are not fine.

Stage A. State (deterministic, already code)

python scripts/build_seasonal_state.py

Writes data/seasonal_state.json and data/seasonal_tape.json. Read the state whole:

Each outlier carries ticker, class, sector, ranks{5,10,21,63,126,252}, side, agree_horizons, path_turn_td, cycle{phase,n,k,mean_atr}, all_years{n,k,mean_atr}, ext{r5_pct,r10_pct,r21_pct}, atr14, close, adv_usd_21d. The cycle and all_years stats are already entry-anchored at T+1 over the smallest agreeing horizon. They are a screen. The check still recomputes them at the entry the idea will actually use.

Read the tape whole and SORT it. Do not look up the names you walked in with.

The lab and the data map

Every check script starts from pitch_lab.py at the repo root:

import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[3]))  # repo root
from pitch_lab import *  # noqa

Conventions (its docstring is the authority): returns are FRACTIONS in and PERCENT out of summarize; entry is signal on close D, MOC on D+lag, lag=1 is the real order; ATR is Wilder-14; controls are own-drift, all-days and local +/-126td; N < 15 is judged by the exact sign_test, not a t-stat. battery() is round 1; horizon_scan and episode_paths are round 3. Do not re-derive any of this in the day folder. A reusable helper gets promoted into pitch_lab.py with a test the same morning.

Seasonal machinery that already exists, and is not rebuilt:

month_window_anchors counts the incomplete current month as a month end: drop the current month from any month-end history. The rest of the data map is the pitch skill's table and applies unchanged.

Stage B. Survey the whole surface, then select from it

Write scratch/seasonal_checks/<YYYY-MM-DD>/00_surface_map.md FIRST. It enumerates the whole surface and gives every cell a verdict. A dismissal needs a reason; a cell with no line is a cell never looked at. No map, no publish: the gate reads the disk on the ideas path and the stand-down path.

B1. Four enumerations, each exhaustive

1. Every rank outlier, by asset class AND by sector. ranks.by_class and ranks.by_sector are the checklist. Every class and every sector with outliers gets a verdict line naming its outliers, their side, how many horizons agree, and CHECK or PASS with a reason. A class with outliers and no verdict line fails the survey. Note path_turn_td where the expected path turns inside 5 sessions: an entry at the turn and an entry today are different trades.

2. Every calendar cell inside 10 sessions, crossed with the ten asset classes. Take every entry in calendar_cells, not just the next one, and cross it with:

You are not checking every cell. You are deciding in writing which deserve a check and why the rest do not. "Turn-of-month x credit: not examined" is legal only when the next words say why. Cycle year is a conditioner on every cell above, not a cell of its own.

3. Every active watchlist entry, with today's value of its trigger number. Trigger moved: CHECK. Unchanged: PASS, still citing today's value. Expired: prune at the post-publish rewrite.

4. Every board ticket, one verdict line each. The board is a candidate feed; a ticket that deserves a check gets one from scratch, and a ticket that does not says why (sector beta on its face, inside a staged book signal, entry already past the window).

B2. Select 8 to 12 candidates from the map

Name the novelty axis on each:

Horizons run 3 to 63 sessions. Instruments: US equities, ETFs and futures (never reject an idea for lacking an ETF wrapper). No options.

Coverage floors. Requirements, not targets.

• 8 to 12 candidates touching at least four asset classes from the table.
• At least one non-equity vehicle and at least one short.
• At least one calendar-cell candidate and at least one rank-outlier candidate. These are different search modes and both get opened.
• Every calendar cell inside 10 sessions and every class with outliers appears in the map with a verdict.

Coverage governs where you LOOK; what ships is a separate question. Standing down after a complete survey is a fine morning; shipping three equity longs without having opened rates, metals or FX is not.

Before spending a check, run the candidate against data/seasonal_agent_negative_registry.md (inlined as negative_registry), and only that registry. A collision does not kill automatically, but the write-up says what is different. Once scoreboard carries graded ideas, read its per-axis and per-grade splits and note the read in the map; while the count is a handful, say so and move on.

Stage C. Falsification (the point of the whole thing)

Hand the candidates to checking agents whose brief is to kill them: two or three in parallel, three or four candidates each, then one red-team pass over the survivors together. Every script goes in scratch/seasonal_checks/<YYYY-MM-DD>/.

Each checker's prompt carries: the candidate block from the surface map verbatim with its axis and cell; the paths to the map and the day folder; the adjacent registry entries (own registry, not the whole file); the import boilerplate and conventions line above; the five seasonal rules below with their worked examples; and the standing brief, your job is to kill this; a survivor is a failure to kill, not a success to celebrate. It returns per candidate a verdict (KILL / SURVIVES / NEAR-MISS), the decisive numbers, its script paths, which substantive kill it was, and for every NEAR-MISS the number it turned on.

Round 1. The pitch battery

The pitch's round 1 unchanged (pitch_lab.battery()): pattern against own-drift and all-days controls, N and worst window, era stability, registry collision, book overlap, cost against edge, volatility events inside the window.

The five seasonal rules. Every check answers all five.

These were written from the 2026-09-30 kills (scratch/pitch_checks/2026-09-30/ sb1, sb2, sb4, sb5). Each worked example is a real number from that morning.

1. Entry-anchored windows only. The historical window starts at the close the order would fill at, never at the day-of-year close and never at lag 0. If the idea waits for a path turn, measure from the turn. Example (sb2_txn.py): the board's TXN short read "midterm 5/6 lower"; measured from the T+5 entry it was 1/6. The whole edge lived in sessions the order could never own.

2. Index and sector residual, with its own hit rate and sign test. Report the return net of SPY and net of the sector ETF, each with hit count and sign_test p. A zero residual means the pattern is the index's or the sector's: pitch it as that seasonal or kill it. Examples: WMT's +1.98% window was XLP's; WMT minus XLP was +0.05%, 3/6. GS was XLF, residual 10/16. And the sector itself gets the same test: XLP minus SPY over its window was -0.02%, 11/26 (sb5_xlp.py).

3. Cycle cells show drop-best, drop-two-best and the non-cycle cohort. A conditioner that adds nothing has to be visible. Example (sb1_trv.py): TRV October in midterm years +7.66%, 5/6; drop-two-best +4.32%, 3/4; non-midterm +4.00%, 16/20. The cycle added size, the pattern stood without it, and removing the best years did not flip the sign: that passes this rule. A cell where drop-best flips the sign is a kill (the registry holds several).

4. Regime branch. Split history by SPY within 2% of its 52-week high at entry versus not, and by the fragility dial where the PIT history allows (state which vintage). Say which branch today sits in. Example (sb4_redteam.py): TRV near-high years +1.10%, 7 up / 5 down, which is its own drift; off-high years +8.05%, 14 up / 0 down. Today SPY was near its high. The whole edge sat in the other branch: killed.

5. Recency. The last 10 years individually, each with its return and max adverse excursion in ATR. A long record with a coin-flip last decade is grade C at best. Example (sb4_redteam.py, sb1_trv_dev_b.py): TRV 2016-2025 went 6/4, sign p 0.61 against its base rate, and the losers carried 5 to 7 ATR of MAE. The 26-year record was not today's trade.

Round 2 is mandatory for round-1 survivors

As the pitch: decluster and cluster_note concentration (top two years, one year); definition neighbours (shift the window start +/-3 sessions, the horizon to its neighbours, the rank threshold); era split pre/post 2018 plus the regime the mechanism implies; gate attribution (run it without the rank or cycle gate; if the gate does not move the result, nothing may be credited to it). A b/c suffix on the same script, or a section inside it.

Round 3. Develop the exit from the shape

A survivor is a pattern; a pitch is a trade. Every composed idea gets a development script (_dev suffix), and its path goes in evidence.dev_script. The publisher requires it.

1. Horizon. pitch_lab.horizon_scan across 1 to 63 sessions. horizon_td and exit.time_td come from this table. If the edge peaks at h=14 and fades by h=21, the idea is 14 sessions.

2. Where the dip sits. pitch_lab.episode_paths on winners and losers. If winners typically draw down before they pay, that decides both the entry (MOC now, or a close-anchored LIMIT k ATR lower, or wait for the path turn) and the stop.

3. Entry form. MOC against a close-anchored LIMIT, as WHOLE variants (fill rate plus conditional stats), never a marginal-fill decomposition.

4. Stop from the adverse-excursion table. Tabulate intraday MAE in ATR per year and the result under 0.8 / 1.0 / 1.3 / 1.6 / 2.0 / 2.5 / 3.0 ATR stops, then choose one of three forms and say why:

   • time-only, sized off a 3.0 ATR catastrophe distance (stop_atr_for_sizing >= 3.0);
   • a real stop at a level the table supports;
   • a trail armed after a stated MFE (exit.trail).

   Never a default 2:1 ticket. Example (sb1_trv_dev_b.py): TRV's 0.8 ATR stop was touched in 19 of 26 years, 14 of them winners, and cut the mean from +4.84% to +1.23%; a 1.6 ATR stop left +0.79%; no stop kept +4.84%; a 3 ATR stop triggered 8/26. The shape asked for time-only with a catastrophe distance, which the board's fixed ticket cannot express.

5. Loser paths. what_kills_it quotes a number from episode_paths ("losers were down 1.5 ATR by day 6 and never recovered; a close below X by then says the window has failed"), never a generic risk.

The red-team pass

One agent over the developed survivors together: leg-return correlation over the hold window (two seasonal longs in one sector are one trade: merge or drop); overlap with the systematic layers in the state (staged signals, event sleeve, the ledger; live broker positions are deliberately absent); cost at the developed entry form; and the strongest single argument against each idea. If that argument would convince you, it is a kill.

Hard requirement. Every delivered idea carries evidence computed fresh this morning and a survived line naming a consideration that could have killed it and did not.

Small N is not a kill

The pitch doctrine applies verbatim, and seasonals need it more than anything: a cycle cell has one observation every four years. A 6-0 record is sign p 0.016. Quote the record, its exact sign_test p and the per-event edge. No idea needs a t-stat; leave t_stat null below 15. "Insufficient N", "not significant" and "t below 2" are illegal kill reasons standing alone (the publisher prints KILL-LINT:). A small-N kill names a substantive failure and quotes its number: no mechanism, a gate that does not filter, definition fragility, sign instability across eras, a residual that is zero (rule 2), the wrong regime branch today (rule 4), a dead last decade (rule 5), or cost.

Multiplicity corrections price a search: charge them to a sweep's best occupant, never to a cell that arrived with a mechanism attached.

Kills are journaled with reasons, one footer line each. Do not quietly shrink this stage to save tokens.

When one or two survive

Ship them, without asking. The empty slots are paid for in a short_slate block, priced exactly as in the pitch: two named kills per empty slot, at least 8 candidates over at least 4 novelty axes and 4 asset classes, a reason of at least 120 characters, and 1 to 3 near-misses each with the number it turned on. Nothing else relaxes. A thin sweep cannot fill this block: go back to stage B.

{
  "asof": "YYYY-MM-DD",
  "ideas": [ ... one or two, full idea schema ... ],
  "short_slate": {
    "reason": "what the rest of the morning looked like and why it came back empty",
    "candidates_considered": 11,
    "axes": ["rank_outlier", "cycle_cell", "calendar_cell", "relative_value"],
    "asset_classes": ["us_large", "rates", "gold_miners", "energy"],
    "closest": [{"title": "Long TRV into late October",
      "decisive": "off-high branch +8.05% 14/0; near-high +1.10% 7/5",
      "why_died": "SPY within 2% of its 52w high today; 2016-2025 6/4, sign p 0.61",
      "script": "scratch/seasonal_checks/YYYY-MM-DD/c1_trv.py"}]
  },
  "killed": [{"title": "...", "reason": "...", "novelty_axis": "..."}]
}

When nothing survives

Ship a stand-down: NO TRADES email led by the near-misses, an empty tab, a stand_down journal record. The publisher enforces, as for the pitch: at least 8 candidates over at least 4 axes, at least 4 asset classes, at least 6 named kills, a reason of at least 120 characters, a checks_dir holding real .py checks and 00_surface_map.md, and 1 to 3 near-misses with their numbers. "Nothing worth trading" is a claim about the whole surface.

The payload is the short-slate shape with "ideas": [], the block keyed stand_down, and "checks_dir": "scratch/seasonal_checks/YYYY-MM-DD" inside it.

Stage D. Compose

Pick the best three by risk and reward over their horizons, or however many survived. Grade honestly:

Most seasonals are B or C: 26 years of one October is N=26, and one cycle phase of it is N=6. A cycle cell with N=6 competes for the C slot on edge size and mechanism, per the small-N doctrine, and it may win the morning. A grade-C evidence line quotes the record and its sign p ("5-1, sign p 0.11") plus the per-event edge.

Sizing

Default 30 bps of risk per idea. The agent may set 15 to 50 bps by conviction, under these rules:

• 40 to 50 bps only for grade A or B with a clean last-10-years record (rule 5), a positive residual against BOTH the index and the sector (rule 2), and today in the favourable regime branch (rule 4).
• 15 to 20 bps for grade C, for any idea in the unfavourable regime branch, and for any idea whose mean edge is under 2x its catastrophe risk.
• Otherwise 30.

stop_atr_for_sizing is the real stop distance when the idea has one (>= 1.0), and the catastrophe distance (>= 3.0) when the exit is time-only or trailed. The idea's thesis or evidence.summary must say why the size is not 30 whenever it is not.

Prose rules

• No em dashes. No "it's not X, it's Y". No AI throat-clearing or filler.
• The thesis is three to five sentences with a variant perception and who is on the other side (the flow behind the calendar: rebalancing, window dressing, tax timing, index events, insurers' cat-season books). "This window has been green" is not a mechanism.
• Evidence is numbers with N, control, residual, regime branch and era note. One table at most.
• what_kills_it is an observation inside the hold, quoting the loser paths.

Schema

Write data/seasonal_agent_ideas.json:

{
  "asof": "YYYY-MM-DD",
  "ideas": [
    {
      "title": "one line, no ticker soup",
      "grade": "A|B|C",
      "novelty_axis": "cycle_cell",
      "horizon_td": 21,
      "legs": [
        {"ticker": "XLE", "side": "LONG", "weight": 1.0},
        {"ticker": "SPY", "side": "SHORT", "weight": 1.0}
      ],
      "entry": {"type": "LIMIT", "anchor": "CLOSE", "atr_mult": -0.5, "fill_window_td": 3},
      "exit": {"time_td": 21, "time_order": "MOC", "target_atr": null, "stop_atr": null,
               "trail": {"arm_atr": 2.0, "trail_atr": 1.5}},
      "sizing": {"mode": "risk_bps", "risk_bps": 30, "stop_atr_for_sizing": 3.0},
      "thesis": "...",
      "evidence": {"summary": "...", "n": 26, "t_stat": null, "window": "2000-2025",
                   "control": "own drift and XLE minus SPY", "era_note": "...",
                   "table": [["cohort","N","avg","hit"], ["...","...","...","..."]],
                   "script": "scratch/seasonal_checks/YYYY-MM-DD/c3_xle.py",
                   "dev_script": "scratch/seasonal_checks/YYYY-MM-DD/c3_xle_dev.py"},
      "survived": "the consideration that could have killed it and did not",
      "what_kills_it": "...",
      "overlap": "what the book or sleeves already hold that correlates, or None",
      "changed_since": "only when re-pitching inside 10 td, or matching a pitch fingerprint"
    }
  ],
  "killed": [{"title": "...", "reason": "...", "novelty_axis": "..."}]
}

trail is optional. Vocabularies, and nothing else is legal:

novelty_axis rank_outlier | cycle_cell | calendar_cell | path_turn |
             relative_value | instrument_translation | inversion |
             historical_analogue
entry.type   MOO | MOC | LIMIT
             LIMIT also needs anchor (OPEN|CLOSE), atr_mult (signed),
             fill_window_td (1..10)
exit         time_td is ALWAYS present (1..horizon_td), time_order MOC|MOO,
             target_atr and stop_atr optional
exit.trail   optional {arm_atr, trail_atr}: trail at trail_atr ATR once MFE
             reaches arm_atr. Makes the idea MANUAL for auto-placement
horizon_td   3..63
sizing.mode  risk_bps (default 30, allowed 15..50) | nav_pct (index or carry
             constructions; say why in the thesis)
             stop_atr_for_sizing >= 1.0; >= 3.0 for time-only or trailed
             exits (the catastrophe distance)
             HARD CAPS (the validator fails the publish): risk_bps <= 100
             per idea, ATR risk <= 60 bps per idea and <= 150 bps across
             the slate, nav_pct <= 0.5, at most 4 legs per idea
legs         side LONG|SHORT; a futures leg adds sec_type "FUT", contract,
             proxy_ticker and multiplier

ATR is Wilder-14 on the traded instrument, never the book's simple mean. Placement follows from the grammar: a CLOSE-anchored LIMIT places with its stop and target; MOO and MOC place with a time exit only, and a price stop or target on them makes the idea manual; an OPEN-anchored LIMIT goes in the post-open pass; any futures leg is manual; any exit.trail is manual.

Background-agent completion gate

Every checker, red-team agent and background task started in this run is joined before publishing: keep their task IDs, wait for each to reach a terminal state, read every result, and fold its evidence or failure into the verdict. A time-limited wait is not completion. If an outer deadline nears, finish with a fully evidenced stand-down; never publish unchecked ideas. The skill returns only after the publish command has finished and its validation result has been read.

Publish

python daily_pitch.py --product seasonal --ideas data/seasonal_agent_ideas.json --validate-only   # check
python daily_pitch.py --product seasonal --ideas data/seasonal_agent_ideas.json                   # ship

The publisher validates, sizes every leg, captures yesterday's Approve cells off the Seasonal Agent tab before overwriting it, sends the Daily Seasonal email, rewrites the tab, and appends to data/seasonal_agent_journal.jsonl. Iterate on --validate-only until it is silent. --dry-run --html-out preview.html renders without sending. Fix validation errors by fixing the idea, never by loosening the grammar.

The gates read the disk, not the payload, exactly as the pitch's do:

• scratch/seasonal_checks/<asof>/ exists and holds 00_surface_map.md plus at least one .py check;
• every evidence.script and evidence.dev_script resolves to a file inside that folder (a path into another day's folder fails as stale);
• every composed idea carries dev_script;
• fewer than three composed ideas needs its short_slate block.

A directed idea (non-empty directed_by, McKinley's wording) skips the survey and dev_script rules and none of the others.

After publishing

1. Append any reusable kill to data/seasonal_agent_negative_registry.md: the cell, the rule it failed (1 to 5 above or a battery item), the number, and the script path.
2. Update data/seasonal_agent_watchlist.json: append today's near-misses and closest entries with title, cell, the trigger number ("SPY closes more than 2% below its 52w high before the window opens"), script path, source and expiry (default 15 td; a dated seasonal may park to its next window). Remove entries that expired or fired today.
3. Leave every check in scratch/seasonal_checks/<date>/ as the audit trail.
4. Grading runs on its own:

python scripts/grade_pitch_journal.py --product seasonal

It replays every idea, approved or declined, against its stated entry, exits and trail, and rewrites data/seasonal_agent_scoreboard.json. That is where the product earns its keep or gets retuned or killed.

Standing down

If the price cache is broken or the rank file is stale (ranks.rank_date not the prior session), the seasonal run does not ship on stale data. Say so and stop. A missed morning delivers nothing; it never delivers stale ideas late in the session.