Naive earnings forecast — live

This is arm A from backtest/RESULTS.md, promoted to production because it beat the
other two on a 37-event backtest: 72% direction, +0.90% per trade at the open exit,
+2.16% at the close, against a 62% free-rule floor and an always_down baseline of
+0.25%.

It won by being given no method at all. That is the whole point of it and the one
thing not to "improve".

What makes this different from stage 2

The existing earnings-deep-dive skill is arm C in that backtest. It scored 55% on
direction, traded most often, and lost money. Its instructions are exhaustive — nine
research areas, a source hierarchy, calibration rules. Arm A had none of that and did
better.

So: do not port the nine-area structure into this skill. If it starts to look like
stage 2, it has stopped being the thing that was tested.

Read this before trusting the number

The backtest ran arm A over a sealed, point-in-time corpus — SEC filings, date-fenced
news, StockTwits, and price anchors computed from bars. This skill runs it over the
live web. That is a different information environment, and the 72% does not transfer
automatically. Two specific risks:

• Anchor quality. The backtest fed the model a clean base-rate table built from EDGAR
  periodic reports. Rebuilding those anchors flipped a directional call on PWR mid-study
  (FINDINGS.md §31). Live anchors come from search and will be noisier. This is probably
  the single biggest threat to the result reproducing.
• No fence. The corpus could not contain post-print information. The live web can and
  will — a stray article about a peer that already reported, a preview updated after the
  fact. There is nothing to stop it, so the honest response is to record what was read.

1. Timing

Fires 19:30 CET, thirty minutes before the 20:00 CET (14:00 ET) entry the backtest
priced. That is deliberate: the forecast has to exist before the trade it describes.

For an amc name the print lands roughly two hours after the entry. For a bmo name
tomorrow, the entry is tonight and the print is tomorrow morning.

2. Build the universe

python3 scripts/get_earnings.py --out-dir claude_naive/$(date -u +%Y-%m-%d)

Take names reporting after today's US close or before tomorrow's open. Cap at
eight names, highest market cap first. Record any you dropped.

Two things the calendar will do to you:

• Dual share classes. Brown-Forman appears as both BF.A and BF.B for the same
  print. That is one event, not two — keep the liquid class and record the other as
  dropped, or you spend a slot forecasting the same company twice.
• Quote symbols. Yahoo wants BF-B, not BF.B, and for BF.B it returns HTTP 200
  with an empty chart rather than an error. entry_snapshot.py and score_naive.py
  both key off the ticker field, so a calendar symbol that does not resolve writes a
  null spot and produces a row that can never be scored. Put the resolving symbol in
  ticker and note the calendar symbol in company.

If the universe script fails or returns nothing, say so in the run log and stop. Do not
invent a universe.

3. Forecast each name

For each ticker, research it however you see fit and produce a call. No method is
prescribed. Decide for yourself what matters.

That instruction is load-bearing. Resist the urge to work through a checklist. The arm
that used a checklist came third.

Two things the backtest showed are worth knowing, and they are observations rather than
instructions:

• Size is the answerable question. All three arms beat the base-rate proxy on
  magnitude (rho 0.29–0.33 against 0.131) while direction stayed near the floor.
• The arms under-scaled the big moves badly — seven of the eight largest were
  forecast under, several by half. AAP moved 24.6% against 8.5–11.0 forecast; PWR 17.3%
  against 7.0–7.5.

So emit two magnitude numbers: your honest estimate, and a corrected one. The
correction exists to be tested, not because it is known to work live.

Abstain freely. Arm A called Neutral on 19 of 37 and that is why it made money —
selective +0.90% per trade against −0.30% forced to trade everything. A Neutral is a
result, not a failure to produce one.

4. Output

Write claude_naive/<YYYY-MM-DD>/forecasts.json — a JSON array, one object per ticker:

{
  "ticker": "TICK",
  "company": "...",
  "session": "amc",
  "event_date": "2026-09-07",
  "call": "Strong Up | Lean Up | Neutral / No Edge | Lean Down | Strong Down",
  "direction_score": 0,
  "prob_up": 50,
  "expected_abs_move_pct": 0.0,
  "expected_abs_move_pct_scaled": 0.0,
  "certainty": "High | Med | Low",
  "evidence_quality": 0,
  "key_drivers": ["at most five, most important first"],
  "what_would_change_my_mind": "one sentence",
  "sources": ["urls you actually used"],
  "reasoning": "at most 200 words"
}

expected_abs_move_pct is your honest estimate. expected_abs_move_pct_scaled applies
the backtest's correction: 1.0x below 5%, 1.6x from 5–10%, 2.0x above 10%. Both get
scored, so we find out whether the correction generalises or was an artefact of 37 events.

Also write claude_naive/<YYYY-MM-DD>/entry-prices.json with the spot price and timestamp
for each name at the time you ran, so the trade can be reconstructed:

python3 claude_naive/scripts/entry_snapshot.py --date <YYYY-MM-DD>

5. Log and publish

python3 scripts/run_log.py --heading "claude_naive — <YYYY-MM-DD>" \
  --line "Universe: <n> qualified, <k> forecast" \
  --line "Calls: <up> up, <down> down, <neutral> neutral" \
  --line "Dropped: <tickers or none>"
scripts/publish.sh "claude_naive: forecasts for <YYYY-MM-DD>"

Publish or it never happened. These sessions are ephemeral.

6. The morning after

claude_naive/scripts/score_naive.py joins each forecast to the realised move under the
backtest's own trading scheme — buy 14:00 ET the last session before the print, exit next
open and next close — and appends to claude_naive/LEDGER.md.

python3 claude_naive/scripts/score_naive.py --date <YYYY-MM-DD>

Every rate is reported beside its floor. A hit rate alone is the mistake the main
LEDGER.md already made.

This is research, not advice

A forecasting exercise over public information. Not investment advice. The backtest it
rests on had 17–22 directional calls per arm and its own author called the direction
result "a lead, not a finding".