MiMiTrends

MiMiTrends is primarily an experimental, recreational mathematics project. It began as a personal response to reading books full of stock-market formulas: an attempt to turn some of those ideas into a visible, testable desktop program. It is not a trading system, investment research, or a promise that a statistically unusual move will continue.

The implementation is original and combines the ideas below rather than reproducing one published strategy. The main mathematical reference shelf behind the formulas used in the program is:

• Ruey S. Tsay, Analysis of Financial Time Series — returns, volatility, time-series behaviour, regression, and statistical diagnostics;
• Perry J. Kaufman, Trading Systems and Methods — momentum, trend efficiency, noise, drawdown, and separating persistent movement from an isolated price jump;
• Ernest P. Chan, Algorithmic Trading: Winning Strategies and Their Rationale — quantitative momentum and mean-reversion ideas, signal construction, backtesting, and realistic validation;
• David Aronson, Evidence-Based Technical Analysis — statistical testing of market rules, data-mining bias, and the need to validate apparent patterns rather than trust chart intuition;
• Marcos López de Prado, Advances in Financial Machine Learning — feature engineering, walk-forward evaluation, sample dependence, probability calibration, and protection against overfitting;
• Robert Kissell, The Science of Algorithmic Trading and Portfolio Management — VWAP, liquidity, volume profiles, transaction costs, and market-microstructure context.

These references inspired the project's use of percentage and logarithmic returns, median/MAD robust Z-scores, relative volume, VWAP distance, regression slope and R², path efficiency, realized volatility, logistic probabilities, Brier scores, and walk-forward evaluation. The exact composite scores, thresholds, gap/reversal rules, and UI classifications are MiMiTrends-specific heuristics and should be treated as experiments, not formulas endorsed by the cited authors.

See Financial and statistical methods for method-level sources and the exact implementation locations they support.

A local-first desktop scanner for fresh US and European market anomalies.

MiMiTrends watches a configurable universe of US and European equities, detects unusual recent price activity, ranks the results, and explains why each instrument was selected. Adaptive selection drives the result set; configured thresholds are guardrails rather than a fixed checklist that leaves the table empty.

The application is written in Kotlin/JVM and JavaFX. It is designed as a cross-platform desktop application for macOS, Windows, and Linux. The current builds and user interface have so far been tested only on macOS. Windows and Linux packaging is implemented, but those distributions should be treated as unverified until they receive platform-specific testing.

MiMiTrends is informational software. It does not place orders, provide investment advice, or guarantee that a detected anomaly will continue.

At a glance

Scanner and signal chart

Freshness is the first sortable column. Pattern rows show their watch priority as a percentage, while Outcome progresses from Collecting to a preliminary Beta probability and finally a validated Model probability. The workspace keeps fresh signals, actionable corridors and recoveries, the explanation panel, minute candles, volume, entry, current price, and EMA 9/21 plus the retrospective Trend 30 overlay visible together. Executed trades can be overlaid without leaving the chart.

Positive watch

Positive watch ranks downside-safety candidates separately from anomaly signals. Confirmation counts, entry quality, estimated safety, current-session candles, EMA 9/21, and the latest 30-candle direction remain visible without presenting the result as investment advice.

Scanner settings

Detection, trend, universe, provider, and appearance controls are grouped in one settings window. Each analytical field includes a short explanation of what it changes.

Broker CSV import

The import action accepts the CSV format documented below. A Scalable Capital transaction export can be used directly, while transactions from any other broker can be converted to the same format. Imported executions can be shown on the corresponding price chart without sending portfolio or transaction data to a remote service.

CSV format for completed trades

The CSV file does not have to come from Scalable Capital. You can create it in a spreadsheet or convert an export from any broker, provided that the resulting file follows these rules:

• encode the file as UTF-8;
• separate columns with semicolons (;), not commas;
• include the exact, case-sensitive header names shown below; their order may differ and additional columns are allowed;
• write one transaction per line;
• use yyyy-MM-dd for dates and HH:mm:ss for times. Times are interpreted in the computer's local time zone at the moment of import;
• use European number formatting: a comma is the decimal separator and a period is the optional thousands separator, for example 15,546 or 3.420,12;
• wrap a text value in double quotes if it contains a semicolon. Represent a double quote inside a quoted value as two double quotes ("").

The required header is:

date;time;status;reference;description;assetType;type;isin;shares;price;amount;fee;tax;currency

Column	Required value and meaning
date	Execution date in yyyy-MM-dd format, for example 2026-08-03.
time	Execution time in HH:mm:ss format, for example 21:22:05.
status	Transaction status. Use Executed for a completed trade. Rows whose status is Cancelled or Cancel (case-insensitive) are ignored.
reference	Broker transaction ID. It should be unique and stable; repeated non-empty references are imported only once. May be empty if the broker provides no ID.
description	Instrument or company name displayed with the transaction. May be empty.
assetType	Asset category, normally Security for a stock or ETF.
type	Trade direction, normally Buy or Sell.
isin	Instrument ISIN used to match the transaction to a chart, for example US83406F1021. It is converted to uppercase.
shares	Executed quantity. Must be zero or positive. Fractional quantities are supported.
price	Execution price per unit. Must be zero or positive.
amount	Total cash amount. Use a negative value for money paid on a purchase and a positive value for proceeds from a sale.
fee	Transaction fee. Must be zero or positive.
tax	Tax amount or adjustment; both positive and negative values are accepted.
currency	ISO currency code such as EUR or USD. It is converted to uppercase.

Empty numeric fields are imported as zero. All supplied numbers must be finite. When reference is empty, MiMiTrends detects duplicate rows from their complete CSV content instead. If the same transaction occurs more than once in one file, the first occurrence is kept.

Example with one purchase and one sale:

date;time;status;reference;description;assetType;type;isin;shares;price;amount;fee;tax;currency
2026-08-03;21:22:05;Executed;buy-220;SoFi Technologies;Security;Buy;US83406F1021;220;15,546;-3.420,12;0,00;0,00;EUR
2026-08-03;21:43:14;Executed;sell-220;SoFi Technologies;Security;Sell;US83406F1021;220;15,508;3.411,76;0,99;-6,53;EUR

Save the file with a .csv extension, click the import button in the MiMiTrends toolbar, and select the file. CSV imports run immediately on a dedicated high-priority worker instead of waiting behind scheduled scans or database maintenance. Matching transactions are stored locally, and the currently displayed chart is refreshed as soon as the import commits so new or changed trade overlays appear without a manual refresh. Importing the same file again does not create another copy of an already stored transaction.

Quick start

git clone https://github.com/senatov/mimiTrends.git
cd mimiTrends
./gradlew run

Open Settings, choose the market universe, and allow the first history bootstrap and scan to complete. Select any table row to inspect its chart. No API key is required for the default data path; Finnhub is an optional enhancement for supported US symbols.

Why it exists

Traditional charts are useful for studying price history, but they often require the user to inspect many instruments manually. MiMiTrends reverses that workflow:

1. collect and normalize recent market data;
2. compare the latest completed candles with the instrument's own historical behavior;
3. reject ordinary movement and unconfirmed volume spikes;
4. rank only fresh, statistically unusual candidates;
5. show the signal, its strength, entry context, volume confirmation, and subsequent movement.

The primary question is not “What did this stock do over the last year?” but “What unusual event is happening now, how strong is it, and is the move being confirmed?”

What the application does

• scans US, European, or combined watchlists;
• refreshes the first HTML page of wallstreetONLINE's performance and most-traded tables before every scan, merges duplicate instruments, sorts the combined discovery set by percentage performance, and evaluates the leading 30 resolvable equities through the same scanner as the configured watchlist;
• detects fresh upward and downward one-minute impulses;
• requires an exceptional price move or candle range, not volume alone;
• confirms candidates using candle structure, relative volume, log-volume anomaly, or immediate continuation;
• applies a configurable minimum absolute move so tiny changes in quiet stocks do not become misleading signals;
• supplements a sparse strict result set with relaxed impulses and persistent rising trends;
• adapts thresholds to retain the strongest defensible candidates instead of treating the configured target as a mandatory quota;
• ranks completed results atomically instead of changing the visible table while a scan is running;
• retains recently published events for up to twenty minutes after they stop qualifying, labels them as Cooling, and decays their ranking score while always giving active signals priority;
• rechecks published Strong and Extreme signals every minute in a separate priority task, updating their rows immediately and stopping when they fall below Strong;
• stores minute OHLCV history, company profiles, derived statistics, scan runs, and signal outcomes in SQLite;
• refreshes visible European quotes and executable bid/ask with timestamped observations from Tradegate, Euronext, Lang & Schwarz, and wallstreetONLINE where an instrument can be resolved safely;
• never labels a crawled quote as fresh using its HTTP download time: the provider's own observation timestamp is required;
• uses exchange-local time zones and market calendars for US, Xetra, Euronext, and Helsinki instruments;
• pauses scanning while every selected market is closed and resumes after the earliest next opening;
• displays quote age in the leading Delay column and sorts it numerically, distinguishing current observations from stale ones without hiding delayed-but-useful European candidates;
• marks statistically repeating two- or three-minute price paths with a bright-red ↻ indicator beside the company name; the tooltip reports the measured cycle strength;
• shortens displayed company names and copied Search Words by removing trailing legal forms, punctuation, share-class labels, and leading or trailing The;
• provides a signal-focused chart with a consistently styled full-history fallback;
• imports completed transactions from compatible broker CSV files and outlines each matching trade interval on the chart with a translucent orange highlighter-style frame, a draggable rounded explanation card, and a curved purple leader that keeps multiple trades identifiable;
• remembers window geometry, divider position, table appearance, columns, and the selected instrument.

Reading the scanner table

The primary table intentionally uses plain-language categories instead of exposing every raw statistical value. It distinguishes active signals from recent context: a Cooling row records a signal that was published recently but no longer passes the current detector. It is not presented as a continuing impulse and expires automatically. A rapid opposite V-reversal updates the same visible episode instead of silently erasing it.

Column	Meaning
Delay	Age of the latest provider observation. It is numeric and sortable; the icon distinguishes a current quote from a stale one.
Symbol	Instrument identity and cleaned cached company name. A bright-red ↻ marks a statistically repeating two- or three-minute price cycle; its tooltip shows cycle strength. The context menu copies either the ticker or the same short search-friendly company name.
Pattern	Detected fresh impulse, relaxed impulse, or persistent trend. A bracketed value such as [40%] is the watch-priority score derived from structure, anomaly strength, entry timing, freshness, volume, and available outcomes. It is a ranking aid, not a predicted return or recommendation.
Move 10m	Signed price change over the latest ten-minute display window.
Price	Latest completed locally available price in the selected display currency.
Anomaly	Composite anomaly ranking: Low, Moderate, High, or Very high. This measures how unusual and well-confirmed the detected move is; it is not a buy/sell recommendation or a prediction that the move will continue.
Outcome	Collecting means that no completed independent episode is available. Beta 60% is a deliberately labelled preliminary, beta-smoothed probability from an incomplete or concentrated sample. With representative history it shows median directional return and the empirically profitable share, such as +0.08% · 58%; an accepted trained model is labelled Model 60%. Tooltips identify the state, horizon, sample count, uncertainty, and model metadata where available.
Price action	Human-readable interpretation such as Rare impulse ↑, Steady trend ↑, or Volatile / unstable. Rare impulses receive a dark-green outline.
Volume	Normal, Elevated, Strong, Extreme, or Price-led, with relative volume when available.
Age	Signal window or retained publication age; this is separate from quote freshness.
Turnover	Approximate current-session traded value.
Updated	Provider observation time rendered in the user's local time zone.

Hovering the derived columns reveals the exact jump, range, volume Z-scores, RVOL, composite score, and an explanation of the classification. This keeps the normal workflow readable without discarding analytical detail.

Statistical model

MiMiTrends evaluates completed, minute-aligned OHLCV bars. Bars with non-finite values, impossible OHLC relationships, negative volume, invalid prices, or malformed timestamps are excluded. A valid zero-volume bar may remain in the price series; positive historical volume is still required when constructing a relative-volume reference.

The scanner also recognizes fresh V-shaped reversals across two- to six-minute shock windows. It requires a statistically unusual shock of at least 0.25%, a recovery of at least 0.20%, directional efficiency on both sides of the turn, two confirming recovery steps, sufficient recovery velocity, and limited lingering near the extreme. Scoring combines time-matched shock rarity, reclaimed depth, fall/recovery speed, path quality, and freshness. The extreme expires after nine minutes, so an unrecovered fall, slow weak bounce, repeatedly tested bottom, or old inactive reversal is not promoted as a current opportunity.

Clean staircase-like growth is ranked separately as Steady rise ↑. The detector evaluates 10- to 180-minute windows and requires a meaningful return, regression fit, directional efficiency, a majority of non-negative minute steps, bounded drawdown, and positive continuation during the latest five minutes. Once an hour of session data exists, the 60-minute return and regression slope must also be positive, preventing a clean short bounce inside an established decline from being mislabeled as sustained growth. Short 10- and 15-minute windows still allow unusually clean opening trends to qualify before an hour of context exists.

A slightly negative three-minute tail does not immediately invalidate an otherwise continuing rise. It is accepted only when the pullback is no larger than 0.05% and remains below the average absolute minute move of the selected trend window. This reduces signal flicker from ordinary noise without retaining a material short-term reversal.

Comparable baseline

For a candidate candle, the scanner prefers historical candles from prior sessions within approximately ±15 minutes of the same exchange-local time. This matters because volatility and volume near an opening auction are not comparable with quiet midday trading.

The default baseline uses five sessions and requires enough samples to avoid unstable estimates. If comparable time-of-day history is still too short, the scanner falls back to a larger recent sample.

Sessions are calculated in the instrument's exchange time zone:

• US listings: America/New_York;
• Xetra and most continental European listings: Central European time;
• Helsinki listings: Europe/Helsinki.

Robust Z-scores

Ordinary mean and standard deviation are easily distorted by the same outliers the scanner is trying to detect. MiMiTrends therefore uses the median and median absolute deviation (MAD).

For observations x:

center = median(x)
robust scale = 1.4826 × median(|x - center|)

A small floor prevents near-zero historical variation from producing infinite or absurd scores.

The impulse model calculates:

• Jump Z — absolute close-to-close return deviation from the robust baseline;
• Range Z — excess high–low range relative to the normal candle range;
• Volume Z — positive anomaly in log(1 + volume);
• RVOL — candidate volume divided by the median positive baseline volume;
• body ratio — absolute candle body divided by its high–low range;
• close location — where the close lies inside the candle range.

These technical values are persisted and remain available in tooltips, while the table converts them into plain-language price-action and volume labels. Volume quality is stored with every minute bar as REPORTED, ZERO, MISSING, or ESTIMATED. Missing, zero, and partial volume remains visible but receives no statistical confirmation and is excluded from volume baselines.

Impulse qualification

A candle is eligible only when all of the following are true:

1. the signal is inside the configured freshness horizon;
2. Jump Z or Range Z crosses its threshold;
3. the absolute price move crosses the economic minimum;
4. the close is directional rather than buried in a long opposing wick;
5. the move is confirmed by volume, body quality, or same-direction continuation.

Volume cannot independently promote a quiet instrument.

The default strict thresholds are:

Jump Z             3.0
Range Z            3.5
Volume Z           2.0
Relative volume    1.8×
Body ratio          0.55
Absolute move       0.20%
Maximum age         2 minutes

All user-facing thresholds can be adjusted in Settings.

Additional public providers

European coverage is corrected by independent provider adapters rather than by overwriting the Yahoo series. Tradegate and Euronext can be enabled and paced in Settings. Lang & Schwarz is a separate, disabled-by-default personal-use quote fallback for resolvable ISINs or unambiguous company names. When explicitly enabled, it reads the Germany, Europe, and Euro Stoxx tables in one bounded pass, accepts only recent timestamped bid/ask snapshots, and never converts their midpoint into an analytical candle. Before every normal scan, wallstreetONLINE discovery reads only the first HTML page of the public top-performance and most-traded tables. It merges duplicate paths, sorts the combined set by reported percentage performance, keeps the first 30 rows, resolves their equity tickers, and adds them to that scan's market universe. This is candidate discovery, not an endorsement and not a replacement for the normal signal quality gates. Separately, the low-priority quote crawler opens at most five matching instrument pages per pass and accepts a quote only when its ISIN matches the canonical instrument ISIN. The site's disallowed stock-search RPC is not used.

Every quote observation is stored in provider_quotes with provider, identifier, MIC, currency, bid/ask, and original observation time. A quote without an unambiguous provider timestamp is rejected. Quote-only sources can refresh a published row, but never create or extend analytical candles and therefore cannot manufacture price patterns or volume confirmation. Only genuine OHLCV history from Yahoo and minute bars aggregated from Finnhub trades may extend the series used by the detectors.

These are public website integrations rather than contracted APIs and can change without notice. Users must enable and use Lang & Schwarz only when their use complies with that website's terms. Failures are isolated per provider, logged without cookies or credentials, and do not stop the rest of the scan.

Early three-minute momentum

The scanner also evaluates the latest three consecutive completed candles as one directional move. This catches a rapid rise or fall before any individual minute crosses the ordinary absolute-move threshold. The three-minute move must exceed 0.35%, remain at least 65% directionally efficient, continue in the latest minute, and be statistically unusual against comparable three-minute windows from prior sessions. Its score uses the same robust price and volume scale as an ordinary impulse. Because only the latest three candles are eligible, the signal disappears automatically when acceleration stops; a flat ten-minute tail cannot remain an active momentum signal.

Range and chop rejection

Fresh impulses, momentum, and reversals are checked against the preceding fifteen-minute price regime. A window is treated as range-bound when net displacement is small relative to the total travelled path and price repeatedly alternates direction. A candidate that remains inside that established high/low range is rejected. A close beyond the range with a small confirmation buffer is retained as a possible breakout. This prevents ordinary oscillation around a mean from being promoted while preserving genuine exits from consolidation.

Repeating short price cycles

The latest 12–24 completed minute bars are also checked for bounded paths that repeat after two or three minutes. The detector requires at least four direction changes across at least nine bars, low net displacement relative to the travelled price path, and autocorrelation of at least 0.60 at lag two or three. A second pass removes the fitted local center before measuring correlation, allowing the same repeating traffic pattern to be recognized while its price band shifts and later returns. A directional rise therefore does not qualify merely because adjacent returns are similar. Detected cycles remain candidates for inspection but are marked with a bright-red ↻ beside the company name; the marker describes observed structure and is not a prediction that the next phase will repeat.

Composite score

The raw impulse score gives most of its weight to the stronger price anomaly, then adds volume confirmation and immediate continuation:

raw = 0.65 × max(Jump Z, 0.8 × Range Z)
    + 0.25 × Volume Z
    + 0.10 × continuation

The result is multiplied by candle quality and exponential freshness decay. Consequently, a structurally clean current impulse ranks above an old or wick-heavy event with similar raw Z-scores.

Relaxed impulses and trend fallback

If strict impulses do not fill the configured minimum result count, MiMiTrends may add:

• relaxed impulses, using moderately lower statistical confirmation thresholds while retaining freshness and the absolute-move floor;
• persistent trends, which do not depend on one exceptional candle.

The trend fallback evaluates up to 180 minutes of the current session. It requires positive net return, a positive least-squares slope, sufficient R², an efficiency ratio above the configured minimum, and continued growth over the latest ten and five minutes.

Trend efficiency is:

net progress / total travelled price path

This allows reasonable pullbacks while rejecting a noisy sideways chart that merely ends slightly above its starting point.

Signal outcomes

Published signals are evaluated later at target horizons of 5, 10, and 30 minutes. Outcome returns use the exact signal candle price and timestamp rather than the later scan-completion price.

The scanner's Outcome column uses independent ten-minute episodes. Repeated publications for the same symbol and signal family within fifteen minutes count as one episode. Directional returns are reduced by a conservative 0.20% friction allowance before profitability is assessed, so a negligible move is not counted as a useful continuation. As soon as at least one completed episode exists, the column exposes a mild beta-smoothed estimate labelled Beta; this early number can change sharply and is not a validated model. A representative empirical distribution and its 95% Wilson interval require at least twelve independent episodes, five symbols, and three trading days. Until then return quartiles and excursion statistics remain hidden rather than being inferred from an inadequate sample.

Calibration is walk-forward: a signal can use only outcomes recorded before its own signal time. When enough history exists, the first estimate is restricted to a comparable cohort by market region, strict versus relaxed detection, and realtime versus non-realtime feed quality. A broader signal-family estimate is used only while that cohort is too small. Detector scores are converted to a 0–10 historical percentile after at least thirty prior observations from the same signal family and direction, making unlike detectors safer to rank in one table.

Adaptive selection uses three explicit quality tiers. Strict candidates remain first, at most two strong adaptive candidates may follow, and the scanner may fill the current list to seven or eight rows with structurally valid lower-confidence candidates marked watch. The watch tier still rejects stale, cooling, and statistically insignificant observations; it is not a blind row quota.

Before publication, every result passes an additional attention-quality gate. Current impulses require a meaningful market-specific move, strong price or range deviation, a directional candle body, and either volume confirmation or exceptional price evidence when volume is unavailable. Signals older than two minutes and cooling context never qualify as current candidates. Trend candidates need at least a 15-minute formation window, 0.35% progress, and sufficient path efficiency. Calibrated adaptive results must rank at or above the 80th historical percentile, and at most two preferred adaptive rows may be published in one scan. Watch rows use a lower score floor and are visibly labelled rather than presented as equivalent recommendations.

The primary tier is intended for roughly thirty-minute long-candidate review. It rejects upward bounces that remain inside an intraday distribution regime: negative session structure, material distance below VWAP, repeated five-minute drops, repeated high-volume red candles, lower highs/lower lows, or excessive drawdown from the session high. A second, lower-priority downside-watch tier retains only genuinely sharp vertical falls (at least 0.75% with a price/range Z-score of 4.5 or more). These downside observations are useful for watching a possible reversal but are always ranked after qualifying long candidates. During open markets, cooling and previously saved rows are no longer used to fill an otherwise empty candidate table.

An Early recovery ↑ · watch detector covers a different setup: a material intraday decline followed by a base, higher lows, positive recent slope, recovery of at least one quarter of the decline, and a controlled pullback from the recovery high. It is deliberately a watch signal because the original decline can resume. Rejected scans now retain a diagnostic such as RECOVERY_TOO_SMALL, NO_HIGHER_LOW, WEAK_LAST_10M, RECOVERY_RETRACE, or STALE_DATA instead of collapsing every failure into NO_CURRENT_SIGNAL.

While an episode is active, observed candle highs and lows update its maximum favorable excursion (MFE) and maximum adverse excursion (MAE). These path metrics describe typical opportunity and drawdown after a signal; older outcomes created before this schema was introduced remain usable but have no excursion values.

The database stores both the target horizon and the actual elapsed time. This avoids silently treating a delayed observation as an exact five-minute result. Outcome collection accepts only a small scheduling delay and records:

• signal entry price;
• observed price;
• target horizon;
• actual elapsed minutes;
• realized percentage return;
• maximum favorable and adverse movement observed before the horizon;
• observation timestamp.

This apparatus is intended for later empirical work: measuring continuation rates, evaluating thresholds, comparing signal classes, and detecting whether an apparently strong score has predictive value. The application does not yet present these records as a backtest or claim a validated trading edge.

Research dataset and walk-forward evaluation

The analytics database also keeps an episode-sampled research dataset independent of the table publication decision. At most one sample per symbol, signal family, and direction is stored within fifteen minutes. Rejected evaluations become long-direction control samples, allowing detector outcomes to be compared with ordinary market observations instead of only with other published signals.

Each sample captures only information available at observation time: trailing 1/3/5/10/30/60-minute returns, recent range and realized volatility, VWAP and session-extreme distances, relative recent volume, trend efficiency, detector score, anomaly metrics, source quality, and acceptance/publication state. Future 5/10/30-minute returns and favorable/adverse excursions are written later to separate outcome records.

AnalyticsRepository.walkForwardResearchReport evaluates historical probability estimates using earlier trading days only. It reports predicted and actual win rates, Brier score, average friction-adjusted return, and sample/day counts by signal family and direction. This is an evaluation foundation, not an assertion of predictive edge; model fitting and user-facing research reports can build on the same point-in-time dataset. The About dialog's Prediction diagnostics action shows these metrics for the 5-, 10-, and 30-minute horizons and can export the complete locale-independent report as CSV without blocking market-data collection. Its Backfill history action replays retained minute bars at fifteen-minute observation intervals. Each detector invocation receives only bars available at that historical instant; 5/10/30-minute outcomes are attached afterward from the same session. Backfill runs on a dedicated worker, is safe to repeat, and leaves the live scanner executor available.

Prediction maintenance is automatic. Thirty seconds after startup, and then every six hours, a dedicated research worker checks historical coverage, fills missing retained history, and trains Smile regularized logistic models for the 5-, 10-, and 30-minute horizons. Training requires at least 300 completed samples across seven trading days. The latest two days are held out chronologically; a candidate becomes active only when its Brier score improves on the smoothed signal-family baseline and its highest-probability quartile does not underperform the validation population. Models are versioned in SQLite, unchanged datasets are skipped, and rejected candidates never replace the active model.

An active model uses detector metrics available identically in historical and live paths. Its validated probability replaces the displayed beta probability, while historical return/excursion statistics remain visible. The UI identifies the logistic source, model id, and training sample count. If no validated model is available, the existing walk-forward beta calibration remains the automatic fallback.

Scan rotation and source diagnostics

The configured universe is supplemented on every pass by up to 30 resolvable leaders from the first wallstreetONLINE performance/most-traded HTML pages. Open-market scans alternate US and European symbols and rotate their starting positions between cycles, so the tail of a static or discovered list does not systematically receive the oldest evaluation. Recently active candidates remain at the front for three full cycles and also retain the dedicated one-minute priority refresh.

The configured scan interval is measured from the start of one pass to the intended start of the next. If a pass takes longer than that interval, the next pass begins after a short safety delay instead of overlapping the running pass. The status line and INFO scan summary report the completed pass duration, oldest analytical bar age, counts by actual last-bar source such as FINNHUB, YAHOO, or a European provider, and the effective countdown to the next pass. These measurements are intended to reveal a real throughput bottleneck before parallel requests or database sharding are introduced.

Market data

Yahoo Finance

Yahoo Finance is the default history and fallback provider and does not require an API key. The application bootstraps recent minute history, then requests and upserts the missing tail. Yahoo's public chart endpoint is not a contracted API and may change without notice.

Corrective provider quotes may refresh the displayed current price and executable bid/ask, but are never treated as minute bars and never extend the series used by the detectors. Freshness validation follows the analytical series, so a current snapshot cannot disguise stale candle history.

European quotes may be delayed. MiMiTrends labels data as live, delayed, Yahoo, or cached rather than assuming that every last bar is current. Open-market recommendations require the analytical candle history to be no more than three minutes old; an isolated newer quote can update the displayed price but cannot make stale analysis current.

European quote correction

Timestamped public observations from Tradegate, Euronext, Lang & Schwarz, Scalable Capital, and wallstreetONLINE can update only the visible quote and executable bid/ask of European instruments. The leading Delay value is calculated from the timestamp of the latest candle used by the detectors, not from the moment MiMiTrends downloaded a page or from a newer isolated quote. Updated separately shows the timestamp of the latest displayed price. A provider that returns an old quote therefore remains visibly stale and cannot displace a newer observation or make stale analysis appear current.

Finnhub

Finnhub is optional. With a user-provided API key, one WebSocket connection subscribes to selected US symbols and aggregates trades into minute OHLCV bars. Finnhub also acts as a company-profile and logo fallback.

Configure the key in Settings → Finnhub live feed, through FINNHUB_API_KEY, or through an ignored project .env file. The final local fallback is:

~/.mimi/trends/finnhub.properties

Never commit credentials. Provider availability and exchange coverage depend on the user's account and licensing.

Market calendar and closed-market behavior

The scheduler uses exchange-local clocks, weekdays, daylight-saving rules, and local holiday calendars. It distinguishes regular US, Xetra, Euronext, and Helsinki sessions.

If none of the selected instruments is currently tradable, MiMiTrends:

1. stops issuing provider scans;
2. restores the most recent published snapshot when available;
3. clearly marks the rows as saved or cached, not live;
4. calculates the earliest next opening among the selected markets;
5. displays the resume time in the user's local time zone;
6. schedules one wake-up shortly after that opening.

The calendar is local and deterministic, so ordinary closed-market behavior does not depend on an external calendar service. Future or exceptional exchange closures can still require a software calendar update.

Signal-focused chart

Selecting a scanner row opens its locally stored candles and volume.

Signal focus is the default mode. It shows a limited context before the signal instead of compressing several days of history into hundreds of unreadable candles. The chart includes:

• full company name and ticker;
• current locally available price;
• signal type, time, entry price, current/exit reference, score, and RVOL;
• a non-linear trading-time axis that removes closed-market gaps and reserves at least one third of the width for fresh signal candles;
• aggregated distant context with unaggregated minute candles around the signal;
• EMA 9 and EMA 21 overlays for short- and medium-speed price direction;
• a dashed, retrospective linear-regression trend across the latest 30 visible candles;
• a highlighted signal interval and signal-volume bar;
• real OHLC, candle return, and volume for the candle nearest the cursor;
• stronger background grid lines for accurate visual comparison;
• a Full history switch for broader context;
• a 1D range that starts at midnight in the instrument's exchange-local time and therefore shows the current trading day rather than a rolling 24-hour mixture of two sessions;
• a Trades overlay with translucent rounded cards, clear borders, curved purple leaders, and preserved drag positions in both chart modes;
• pan, zoom, synchronized time/price cursors, and tooltips.

Holding the pointer over a scanner row for five seconds performs one bounded focused refresh while the market is open. The row's Delay cell shows an hourglass during the request. The refreshed signal, statistics, provider tail, and chart are applied only to the currently relevant symbol, and a per-symbol cooldown prevents repeated provider polling while the pointer remains nearby.

Switching instruments resets all chart ranges after datasets, signal markers, and trade overlays are installed. This prevents a high-priced instrument from leaving an unusable price scale on the next chart.

The chart deliberately does not connect separate trading sessions with a continuous close-price line, because such diagonals can suggest price movement during periods when no trading occurred.

SQLite database and analytics

All durable state is stored locally in:

~/.mimi/trends/mimitrends.db

SQLite runs in WAL mode with foreign keys enabled, a busy timeout, batched prepared statements, and transactional schema migrations. It is a good fit for the application's local, append/upsert-heavy workload and avoids the lifecycle overhead of an ORM.

MarketRepository owns operational minute bars and company profiles. AnalyticsRepository owns derived and historical analytics. Both use the same database but keep their responsibilities explicit.

Operational tables

Table	Purpose
minute_bars	Unique minute OHLCV rows keyed by symbol and timestamp.
company_profiles	Company name, exchange, logo URL, and cached image.

Minute bars use UPSERT semantics, allowing an incomplete or repeated provider candle to be corrected without creating duplicates.

Analytical tables

Table	Purpose
schema_migrations	Transactionally applied schema versions.
instrument_metadata	Name, exchange, currency, time zone, aliases, and tradability.
corporate_actions	Splits and dividends available for validation and future normalization.
market_calendar_rules	Regular exchange hours and time zones.
trading_sessions	Observed session boundaries, bar coverage, volume, and turnover.
fx_rates	Dated currency conversion rates.
data_quality	Feed source, freshness, bar count, status, and diagnostics.
aggregate_bars	Locally produced 5-, 15-, and 60-minute OHLCV.
baseline_stats	Median/MAD return and log-volume by instrument and local time of day.
scan_runs	One durable record for every scanner pass.
scan_candidates	Accepted/rejected symbols, raw metrics, exact signal time and entry price, source, and publication state.
signal_outcomes	Target and actual horizons, observed price, and realized return.
provider_instruments	Persisted provider-specific symbol-to-ISIN mappings, MIC, currency, and resolved name.
provider_minute_bars	Venue-specific observations with provider, ISIN, MIC, currency, and monotonic observation time.
provider_quotes	Latest provider bid/ask, sizes, session totals, average, executions, range, and previous close.
broker_transactions	Locally imported broker executions used for chart overlays and trade context.

Candidate publication and scan completion occur in one transaction. This prevents a partially completed scan from appearing published. Foreign keys and cascading retention protect referential consistency.

Retention

• raw minute bars: 90 days;
• 5/15/60-minute aggregates: 730 days;
• scan runs, candidates, and quality records: 180 days;
• instrument metadata, corporate actions, observed sessions, FX rates, baselines, and outcomes: retained.

More detail is available in Analytics storage.

Project structure

core/         Domain models, market time-zone resolution, OHLCV validation, logging tags
database/     SQLite cache, analytics repository, migrations, aggregation, outcome storage
market-data/  Yahoo and provider-specific HTTP clients, parsers, and response mapping
scanner/      Impulse/trend statistics, market calendar, scanner settings
charts/       Reusable JFreeChart-FX candlestick and volume component
finnhub-ws/   Optional Finnhub WebSocket client, profiles, and minute aggregation
app/          JavaFX lifecycle, dialogs, presentation, state, and orchestration
Doc/          Architecture notes, packaging instructions, and documentation assets
Scripts/      Native packaging helpers

The dependency direction keeps UI code out of the data, scanner, and chart modules. Statistical and persistence behavior can therefore be tested without starting JavaFX.

Requirements

• Git;
• a JDK compatible with the configured Gradle toolchain, or permission for Gradle/Foojay to resolve it;
• network access for first-time dependency resolution and provider requests.

The Gradle wrapper is included.

Build and run

macOS or Linux:

./gradlew run
./gradlew test
./gradlew check
./gradlew build

Windows:

gradlew.bat run
gradlew.bat test
gradlew.bat check
gradlew.bat build

In IntelliJ IDEA, import the repository root as a Gradle project and use the shared MiMiTrends [run] configuration. The application entry point is:

org.senatov.mimitrends.LauncherKt

Using the application

1. Start MiMiTrends.
2. Open Settings and select the US, European, or combined region.
3. Adjust the watchlist, scan interval, freshness, liquidity, and statistical thresholds if needed.
4. Optionally configure a Finnhub API key.
5. Wait for the initial historical bootstrap and completed scan.
6. Sort by Delay to inspect freshness, or by anomaly, price action, volume, and ten-minute move.
7. Hover derived cells to inspect the exact statistical values.
8. Hold the pointer over a row for five seconds to request a focused online refresh; the Delay cell shows an hourglass until it completes.
9. Select a row to open its signal-focused chart.
10. Move the cursor across the chart to inspect candle OHLCV values, or drag a trade card to uncover candles.
11. Switch to Full history when broader context is needed; trade-card styling and positions are preserved.

Settings and UI state are stored beside the database under:

~/.mimi/trends/

Native desktop packages

Native packages contain a private runtime, so end users do not need to install a JDK. jpackage must run on the target operating system; the project does not cross-build native formats.

macOS

Run the packaging script from the repository root to create a complete self-contained DMG. The package includes MiMiTrends and its private Java runtime, so the destination Mac does not need a separate JDK installation.

./Scripts/build-macos-dmg.zsh

The script requires Xcode Command Line Tools and a Developer ID Application certificate in the login Keychain. It automatically selects the first matching certificate, signs the application and embedded native libraries, builds the DMG, notarizes it with Apple, staples and validates the ticket, and prints its location and SHA-256 hash. Before every DMG attempt it atomically increments the three-digit application patch version in gradle.properties. Patch values run from 000 through 099; decimal carry is intentional: 2.65.099 becomes 2.66.000. The same value is then used for the title bar, About dialog, JAR manifest, generated build metadata, package metadata, and DMG filename. Use an explicit identity when more than one suitable certificate is installed:

./Scripts/build-macos-dmg.zsh --identity "Your Name (TEAMID)"

For a public release, first store App Store Connect credentials in a Keychain profile as described in Native packaging, then run the complete notarization workflow:

./Scripts/build-macos-dmg.zsh

Notarization, ticket stapling, and validation are mandatory defaults for distributable DMGs and GitHub releases. Use --no-notarize only for an explicitly local diagnostic package.

GitHub releases include categorized notes generated from commits since the previous release. Start user-visible commit subjects with NEW:, FIX:, CHANGE:, or DOCS: and follow the prefix with a specific description of the observable change. Unprefixed commits remain visible under Other changes, but categorized messages produce a clearer release history.

The notarized form uses the MiMiNotary keychain profile by default; override it with --profile. Apple explains the distribution requirements in Notarizing macOS software before distribution and its current command-line workflow in TN3147: Migrating to the latest notarization tool.

For a local unsigned .app image instead of a distributable DMG, run:

./gradlew :app:packageMacApp

Windows

Run the following command from a Windows Command Prompt:

Scripts\build-windows-exe.bat

The script creates a self-contained EXE installer and prints its location and SHA-256 hash. It requires a JDK, WiX Toolset 3.x, and the WiX candle.exe and light.exe commands on PATH. For public distribution, sign and timestamp the resulting installer with a trusted code-signing certificate and verify its signature. See Microsoft's SignTool documentation and Oracle's jpackage packaging prerequisites.

This Windows packaging workflow has not been tested because I do not have a Windows system 😞

Linux

Run the packaging script on Linux:

./Scripts/build-linux-packages.sh
./Scripts/build-linux-packages.sh --portable-only
./Scripts/build-linux-packages.sh --deb-only

The default command creates both a portable self-contained .tar.gz archive and a Debian/Ubuntu .deb package. Use either option to build only one format. A JDK and sha256sum are required; building the Debian package additionally requires fakeroot and dpkg-deb. Inspect the .deb with dpkg-deb --info and dpkg-deb --contents, then install it on a disposable test system before distribution. See Oracle's jpackage packaging prerequisites and the official Debian documentation for dpkg-deb.

This Linux packaging workflow has not been tested because I do not have a Linux system 😞

Outputs are written below:

app/build/distributions/native/

See Native packaging for signing, notarization, prerequisites, and output details.

Logging and diagnostics

MiMiTrends uses SLF4J and Log4j2. Logs are written to the run console and the rolling file:

/tmp/MiMiTrends.log

The default INFO level records lifecycle events, scan summaries, warnings, and failures without per-tick noise. Enable temporary detail with:

JAVA_TOOL_OPTIONS="-Dmimitrends.logLevel=DEBUG" ./gradlew run

The status bar reports the current operation. Completed scans additionally show pass duration, oldest data age, source coverage, and the effective delay before the next pass. A red details button opens the complete error report. Credentials are not intentionally logged.

Technology

• Kotlin/JVM;
• JavaFX 26;
• JFreeChart and JFreeChart-FX;
• SQLite JDBC;
• Gradle Kotlin DSL;
• SLF4J and Log4j2;
• JUnit 5 and Kotlin Test;
• jpackage for native desktop distributions.

Platform status

Platform	Packaging	Current validation status
macOS	.app, signed/notarized DMG	Implemented and tested.
Windows	Self-contained EXE installer	Implemented, not yet tested locally 😞
Linux	Portable archive and Debian package	Implemented, not yet tested locally 😞

The codebase and dependency selection are cross-platform, but platform-specific testing is still required before claiming production support outside macOS.

License and third-party components

JFreeChart and JFreeChart-FX are LGPL libraries. Review their upstream license terms when redistributing the application. Provider data remains subject to the respective provider and exchange terms.