Opportunity Finder — Method Statement

How the Opportunity Finder decides — every step from the universe to the names in your approval email, with every threshold and formula stated exactly.

The principle: buy weakness only inside strength — never buy weakness inside breakdown. A stock must first prove it is a strong, quality business. Only after it qualifies may the system consider buying it, and only on a controlled dip — never because it fell. Rules decide; AI only explains and classifies news. Missing data is never guessed: a symbol with an unknown input is excluded and flagged.

The funnel at a glance

Step 1 — the universe: which stocks are considered

Data: daily/weekly prices and fundamentals from market data feeds; tradability overlay from the broker (a name must actually be buyable).

Step 2 — qualification: proving strength (score ≥ 80/100, all four buckets must pass)

A stock enters the watchlist only if it reaches at least 80 of 100 points and passes every required test in all four buckets. Failing any single required test rejects the stock regardless of score.

2a · Trend — 35 points, all required

Revised 2026-08-12. Previously three price-vs-average tests — close above the 50-week, 200-day and 50-day averages. The first two measured nearly the same thing, and the daily 50-day test disqualified a stock on exactly the dip the strategy wants to buy. The revision tests the structure of the trend instead of today's snapshot.

The stacked-averages test says the uptrend has been in place long enough to reorder the averages themselves — a two-week bounce cannot fake it, and an ordinary pullback cannot break it, because a dip moves the price, not the averages. A stock must be trading above its 50-day average on the day it enters the watchlist (qualify on strength); after that, dipping through the 50-day is tolerated — often that is the buyable dip — and only a close below the 200-day average, or the ladder un-stacking, removes it.

2b · Business quality — 25 points, all three required

If any fundamental is unavailable, the stock is neither passed nor failed — it is marked DATA_INCOMPLETE and excluded until the data exists.

2c · Relative strength — 20 points, required

2d · Momentum — 20 points, at least one must fire (minimum 6 points)

Step 3 — the lifecycle: the system remembers each stock, day over day

Each symbol carries a persistent state, day over day: candidate → qualified → waiting_for_pullback → buy_opportunity → manual_approval_required, with rejected, disqualified and no_buy as the downward paths. A stock that qualifies but hasn't dipped simply waits — being good is never, by itself, a reason to buy.

Ongoing disqualification (a qualified stock is demoted the day it breaks): the average ladder un-stacks (20-day, 50-day, 200-day, 50-week no longer in order) → disqualified; close ≤ 200-day MA → disqualified; 63-day return falls behind SPY → back to review. A dip through the 50-day average alone no longer disqualifies (revised 2026-08-12) — that is often precisely the dip worth buying.

Step 4 — buy gates: when a dip becomes a signal

Every one of these must hold on the evaluation day:

Position sizing — scales with the dip and with the account (computed against the execution book's live available cash, read fresh each run — never a hardcoded amount):

Hard caps (smaller number always wins): max 10% of available cash per single purchase and per total position in one stock; absolute ceilings USD 2,000 per purchase and USD 3,000 total per stock. Insufficient cash → no buy.

Per-day deployment budget: at most 3 new buys per daily brief, and their sum may not exceed 30% of available cash. Signals are ranked by dip depth; those over budget are deferred (they re-signal on the next dip). This operationalizes "use cash gradually."

Step 5 — the shortlist: drop quality, not drop depth

"Do not select a stock only because it dropped" is written into the strategy file. Every buy signal that survives the gates is graded on why it dropped, using the day's price context, up to 12 news headlines from the last 3 days, and a small language model (gpt-5-mini) that classifies the drop reason and red flags — classification only; all scoring is deterministic.

Where the AI sits — and where it does not

One small language model has one narrow job: read the day's headlines the way a human analyst would, and sort what it finds into fixed baskets. It writes no scores, applies no penalties, and selects nothing — its baskets are inputs to the fixed rules on this page, nothing more.

Drop classification (daily-change thresholds): market-wide if the market fell ≥ 1.5%; sector-wide if the sector fell ≥ 2.0%; company-specific if the stock fell ≥ 1.5 percentage points below both. These price thresholds stand on their own — they are how the system classifies even with zero headlines.

Hard rejects and heavy penalties:

Quality scoring thresholds: revenue growth good ≥ 5% · EPS growth good ≥ 5% · debt/equity low ≤ 80% · cash strong ≥ 4% of market cap · moat proxy: gross margin ≥ 40% or operating margin ≥ 20% · PEG reasonable ≤ 2.0 · margin-of-safety pullback ≥ 8% · RSI 25–45 = healthy oversold (below 25 = panic collapse) · volume ≥ 3.0× the 20-day average = panic selling, ≤ 1.5× = orderly.

Penalties: missing data −5 · unclear news −10 · guidance cut −15 · high debt + weak cash flow −15 · margin decline −10 · overvaluation −10 · concentration −10.

Selection: final score ≥ 80 = strong review candidate · ≥ 70 = moderate review candidate · ≥ 60 = watch only. The brief presents the top 3 names with score ≥ 70. Selection is a quality bar, not a "best of a bad day" ranking — on a day with no quality dips, fewer than 3 (or zero) are selected.

Step 6 — selection to verdict, and what happens after

The selected names appear in the daily opportunity brief and on Decisions with their full reasoning: qualification score, dip depth and tier, drop classification, shortlist score, and suggested amount. Nothing executes without your explicit approval. Approved buys execute only on the dedicated agent-managed book.

After a buy, the position hands off to the Control Tower, which owns the exit: its risk ladder plus a 20-trading-day maximum hold per lot (the time-stop; backtest optimum), after which an exit is proposed — again for approval, never automatic.

Governance and honesty notes (the part most method statements omit)

1. One disabled rule, disclosed. An overextension guard (no buys > 60% above the 200-day MA, or > 100% 63-day return, or RSI > 80) exists in the configuration but is disabled since 2026-06-27: a two-year backtest showed it was the single largest edge-killer (cutting the 20-day edge from +3.14pp to +0.75pp) and performed worst in down markets. Its thresholds are retained so it can be re-enabled if a true-bear backtest ever justifies it. Crash protection lives instead at the portfolio regime gate and the defensive engine's ladder.
2. Calibration is evidenced, dated, and versioned. The 2% minimum dip and the %-of-cash sizing tiers were set by backtest on 2026-06-27 (edge rises monotonically with dip depth in both market regimes; ~5–10% per name, diversified, is the stable sizing zone). The tested window contains no sustained bear market — stated, not hidden.
3. Every evaluation is logged — every symbol, every day, with inputs, score, state transition, decision, and rejection reason — supporting audit and historical replay with no look-ahead bias.
4. No guessing, ever. Any required input that is missing produces DATA_INCOMPLETE and exclusion — never a default pass or fail.
5. Change control. Parameters change only by deliberate deployment of the configuration file; each run's rules_version hash proves which rules were in force.

Version history — earlier rules stay readable

When the rules change, the old version is not overwritten — it is frozen, so every recorded decision keeps the exact rules that judged it.

• Current — Trend tests revised 12 August 2026: the stacked-averages ladder replaced the three price-vs-average tests (the note in step 2a states why).
• As of 5 August 2026 — the version that judged the CSX case file.

Read on

See the method applied to one real day: the CSX case file → The same exact treatment for the Control Tower's trims and sells is coming next.