User Handbook

MarketTrace User Guide

MarketTrace pulls structured events out of official filings, measures how those events actually moved the market, and orders instruments by which deserve a look first. This guide explains when and how to read each screen.

One thing up front — MarketTrace is a review-and-caution tool, not a buy recommender. Every statistically validated signal so far points negative, so the system never asserts a stock will rise. Every screen exists to help you decide whether an instrument or event needs careful scrutiny.

Start here

Five core concepts

Understand these five and every screen becomes readable. Throughout the app, hovering a number, header, or badge shows the same explanation in place.

① Event: One structured event an LLM extracted from a filing (SEC EDGAR / OpenDART). Each carries a direction (positive/negative/neutral), type, confidence, surprise, novelty, and source reliability. The LLM only extracts; all return math lives in separate numeric modules.
② Abnormal return: The instrument's actual return minus the market (index) return over the same window. It strips out the overall market move so you see only the effect of this event. Measured at D+1 / 5 / 20 / 60 trading days after the filing.
③ Validated signal: Only event types with enough samples (n≥5) and a mean distinguishable from zero (p<0.05) count as validated. Every validated signal so far is negative — there is no bullish signal yet.
④ Model vs history: The model's (LLM's) read compared against the event type's validated historical direction. Match = Confirmed, opposite = Conflict, history exists but the model is neutral = Signal.
⑤ Drift: The average abnormal return following a given event type. Rankings, drops, and instrument leans use this drift weighted by confidence and recency.

Color & badge conventions

The whole app speaks one color language. Color alone tells you direction and how much caution is warranted.

Direction (which way the event points)

positivenegativeneutral

Validated-signal badge

ConflictConfirmedSignal

Conflict = model direction is opposite to validated history (look again). Confirmed = match. Signal = a significant history exists but the model read neutral. No badge means there is no validated signal for this type yet.

Drop diagnosis

Persistent riskUnexplained dropPossible overreaction

‘Possible overreaction’ is blue, not green — it marks a rebound candidate, not an assured rise, and the color says so on purpose.

Everyday screens

Search — find instruments · analyze new ones

Find a stock by ticker, name, or alias. If a company isn't in the data yet, you can run filing analysis from just its name.

Typing searches automatically. Korea recognizes a numeric code (e.g. 005930); the US recognizes a ticker (e.g. AAPL).
Click a result's ticker for the instrument timeline; the count on the right shows whether it has data.
If there's no match or you want a fresh run, use Analyze filings with market/ticker/name to start extraction. New events appear in Events when it finishes.

Login required. Running analysis needs an account. Private companies such as SpaceX are not supported.

Picks — what to look at first

Sharp-drop stocks ordered by which deserve review first, combining recent price moves, recent company events, and how similar past events played out. Split into domestic and overseas.

Review firstNeeds checkingLow priority

Each stock spells out why it appears (drawdown, price freshness, recent events, factors) in plain sentences — not picked on price alone.
Rule: if any stock is down 15%+ from its recent high, show all of them; otherwise show the 5 largest relative drops.

This list is not a promise that a stock will rise. It is a review order. Read each reason and check fresh news / filings before deciding.

Events — the list and the detail

Every event grouped by company. This is the system's central screen.

Pick a market with the Domestic (KR) / Overseas (US) tabs, and filter by ticker/company/type/direction in the search box.
Signal filters: All / Conflicts only / Needs review / Validated only — ‘Needs review’ is the queue of conflicts a human hasn't handled.
Each row: type · direction · signal badge · confidence % · date. Hover the % or the badge for its meaning. ‘✓ Reviewed’ marks human-checked events.

Validated-signal card: How this event type has historically moved the market (per-horizon mean abnormal return, t-stat, p-value, sample size) and whether that agrees or conflicts with the model's read.
Review & correct: (When logged in) fix the model's direction/type/company/confidence. Direction/type edits hit the stats immediately; a company change refetches prices and recomputes returns.
Score components: Confidence, surprise, novelty, source reliability as 0–100% bars. The ? next to the title defines each.
Abnormal-return chart: Abnormal / raw / market / sector-adjusted returns per horizon. If empty, the horizons haven't elapsed yet or price data hasn't been collected (not an error).
Evidence · industries · channels: The sentences the LLM cited as its basis, plus related industry and distribution-channel tags.

Rankings — comparing instruments

Every instrument sorted by its confidence- and recency-weighted validated drift — a cross-stock view. The most cautionary names sit at the top (recent, high-confidence events count more; half-life 180 days).

Lean: Overall direction from validated history — Caution · Favorable · Mixed.
Weighted drift: Post-event average returns of validated events, summed weighted by confidence and recency. This is the sort key.
Simple mean: The same events averaged with no weighting. Compare to the weighted drift to see how much recency/confidence changed things.
Validated: Count of events carrying a significant validated signal (n≥5, p<0.05).
Conflicts: Events where the model's direction is opposite to history. The number in parentheses is how many are still unreviewed.
Top factor: The event type contributing most to the weighted drift, with its average post-event return.

Drops — drawdown screener

Stocks down 15%+ from their 20-day high, diagnosed against validated event history.

Persistent risk: Recent event(s) + a validated-negative lean — the fall fits this name's history; caution likely continues.
Unexplained drop: No recent event in the data explains the fall. An observation, not a signal — find the cause first.
Possible overreaction: Recent event(s) but no validated-negative basis. A rebound candidate pending backtest validation (not a buy call).

Rebound backtest panel: “does buying after such a drop pay off?” — fixed rule, out-of-sample, net of costs. Stays an ‘observation’ when data is thin.
Include stale prices checkbox brings in instruments whose prices aren't fresh (marked ‘stale’).

Your instruments

Watchlist · Alerts

Keep the instruments you care about, and get notified when something notable happens.

Add with the ☆ button on the Events list, an instrument page, or Rankings (login required).
Two alert kinds: Conflict (model direction conflicts with validated drift) · Significant (a validated significant event type). Check the bell in the header and ‘Mark all read’.

Evidence & validation

Stats — why you can trust the system

This screen answers “can I trust this tool?” It shows the performance of the methodology as a whole, not a single stock.

Reaction stats: Heatmap of mean abnormal return by event type × horizon (green up, red down, n = samples). Click a cell to expand the actual filings behind that average.
Validated signals: Only types passing n≥5, p<0.05. All negative so far — caution/avoid rather than buy.
Walk-forward backtest: Out-of-sample performance per horizon (net of costs, look-ahead blocked). Compare signal models: event-type history / significant only / macro regime / price momentum / combined / LLM direction.
Confidence calibration: “Does a 0.7 confidence actually hit ~70%?”, binned by band — over/under-confidence, ECE and Brier score.
Macro regime decomposition: The macro signal backtested per series — if the edge concentrates in one series it's real macro content; if all look alike it's just a calendar proxy.

Macro — why it's here

How far economic releases (inflation, jobs, rates…) landed from expectations. Not individual stocks or events, but the regime backdrop they trade against.

What shifts the overall market mood isn't the raw figure but how it differs from expectations (the surprise) — an in-line print is already priced in and moves nothing; only a miss moves the market. So it lets you read drops and events against “is the backdrop favorable or risk-off right now?” The Stats tab's macro-regime backtest actually uses this data as a signal.

Released: The actual figure that was published.
Expected: What the market anticipated. Uses consensus (pooled expert forecast) when available, otherwise a trend baseline.
Surprise (σ): (released − expected) ÷ historical volatility. Positive = better/higher than expected, negative = worse/lower. The σ unit means “how many times the usual wobble”, so indicators with different units compare on one scale.

Also here

Ledger · Passbook

Personal-finance utilities, separate from the stock analysis (login required).

Ledger: upload a password-protected card-statement PDF to organize spend by category, with top categories and biggest spends charted.
Passbook: upload a bank-transaction PDF to organize in/out flows. Keyword rules can auto-recategorize entries.

Admin

Visible to admin accounts only. Manage users and per-account tab visibility — control which tabs each account can see.

Putting it together

Suggested workflows

Chain the screens like this, depending on your goal.

Morning market check

1Scan Picks from ‘Review first’ down, reading the reason sentences.
2Use Drops to sort by diagnosis (persistent risk / unexplained / overreaction).
3Open event detail for anything that catches your eye to check the validated signal and evidence.
4Add names worth following to your ☆ Watchlist for alerts.

Investigating a specific stock

1Find it by ticker/name in Search.
2If there's no data, run Analyze filings to create events (login).
3Read its upside/downside factors and lean on the instrument timeline.
4If the model is wrong, fix it with Review & correct — it feeds the stats.

Deciding whether to trust a signal

1Check Stats → Validated signals to confirm the event type is significant.
2Check the backtest for out-of-sample performance (net return, hit rate, IC).
3Check calibration to see if the model's confidence matches actual accuracy.

Limits & cautions

Use it with eyes open. Here is what this tool does not do.

There is no bullish signal. Every validated signal is negative — no screen asserts a rise.
It is not a price prediction. It's the average move after past events (drift), with no guarantee the next one repeats.
Recent events may show empty returns. They compute only after the D+N windows elapse.
Check price freshness. Values tagged ‘stale’ need refreshing before you act.
Delisting / survivorship bias isn't modeled yet. The data leans toward survivors.
The final call is yours. Confirm fresh news and filings and decide for yourself.