System Architecture

The system operates as a sequential six-stage pipeline. Each stage produces a clean output that feeds directly into the next. No stage is skipped, no manual override exists between stages.

The universe consists of 738 US-listed stocks and ETFs available on Pluang — an Indonesian retail investment platform. Pluang was selected as the universe source because it offers the broadest selection of US equities accessible to Indonesian retail investors among all available local brokers. Using Pluang's listed universe ensures that every ticker flagged by the system is actually tradable, not just theoretically screened from a raw market database.

The universe spans multiple sectors and asset types including stocks, ETFs, and funds — covering Technology, Financial Services, Healthcare, Energy, Communication Services, Consumer sectors, and more.

Each ticker in the universe is displayed as a card showing the key identity and price data needed to recognize and contextualize a stock at a glance.

Three tools are provided to navigate the full 738-ticker universe without manually scrolling through every card.

Four raw market data points are pulled daily for every ticker in the universe. These serve as the direct inputs to the radar score calculation.

The three momentum dimensions are multiplied together into a single composite score. The multiplicative structure means all three conditions must be simultaneously active for the score to be meaningfully elevated — a single strong signal is not enough.

Because RADAR_SCORE is not normalized across tickers, each ticker is benchmarked against its own historical score distribution. The system runs a mini-backtest covering the past 3 years of daily scores per ticker, then extracts the 95th percentile value as the threshold — meaning only the top 5% of historical signal days qualify as genuinely exceptional.

The threshold rolls forward daily: every time the screener runs, one new day is added to the 3-year window and the percentile is recalculated. This keeps the bar adaptive to each ticker's evolving volatility and momentum regime.

Once both RADAR_SCORE and THRESHOLD are computed, distance measures how far today's score sits relative to the threshold — expressed as a percentage. This converts ticker-specific raw scores into a universal ranking dimension.

Only tickers with DISTANCE_% above zero pass through as candidates. The table is ranked by DISTANCE_% in descending order — the higher the distance, the stronger the current signal relative to that ticker's own historical bar.

Z-Score is a supplementary context indicator. Unlike the threshold (which is derived from only the top 5% of days), Z-Score is computed against the full 3-year distribution — all trading days, not just signal days. It answers the question: how unusual is today's radar score relative to the ticker's average behavior?

Z-Score does not affect ranking or Sniper candidacy. It is displayed as a manual sanity check: a very high DISTANCE_% with a low Z-Score may indicate the threshold itself has shifted (e.g. due to a volatile historical window), not necessarily that the signal is fundamentally stronger.

The final output is a full table of all tickers in the universe, sorted by DISTANCE_% from highest to lowest. Each row shows the complete radar reading for that ticker on the current day — including those with negative distance, which are still displayed for reference. Only tickers with DISTANCE_% above zero are forwarded as candidates to the Sniper Validator for deeper analysis.

For each candidate ticker, the engine replays the past 3 years of daily data and identifies every day on which the Radar score exceeded that ticker's rolling threshold. Each such day is recorded as a Signal Day — a historical moment where the system would have triggered an entry. From each signal day, forward returns are measured across three horizons to simulate actual trade outcomes.

Across all signal days, six aggregate metrics are computed per ticker — three win rates and three average returns, one for each forward horizon. These six values are the foundational inputs to the entire edge and quality pipeline.

Edge measures the risk-adjusted expected value of the signal across all three time horizons. Each horizon produces its own edge value by weighting average return by win rate, then the three are averaged into a single raw score. A positive edge means the signal has historically produced a net positive outcome; zero or below means it has not.

The tanh normalization compresses extreme outliers without distorting the distribution of typical values. A ticker with an unusually massive average return won't dominate the ranking at the expense of more consistent performers. The normalization is applied only internally for scoring — the table displays the raw pre-tanh value for transparency.

Each ticker is assigned a behavioral label based on the directional pattern of its per-horizon edge values. This describes how the profitability of the signal evolves over time — whether it peaks early, compounds late, or behaves inconsistently.

A signal with strong edge metrics but only 2 or 3 historical occurrences is statistically meaningless — the results could be pure luck. Sample Score rewards tickers with more signal days using a square-root scale to reflect diminishing returns: going from 2 to 10 signals matters far more than going from 40 to 50.

A signal that fired 20 times but all within the same 3-week bull run is regime-dependent — it only works in one specific market condition. Cluster Score detects this by grouping consecutive signal days into clusters (gap below 5 trading days = same cluster) and measuring what fraction of signals represent distinct, independent episodes.

A score near 1.0 means almost every signal occurred in a separate market episode — the strategy has worked across many different regimes. A score near 0.0 means nearly all signals were clustered together, suggesting regime dependency.

Even with good average returns and wide cluster distribution, a signal can still be unreliable if its returns are wildly inconsistent — one massive win skewing an otherwise poor record. Stability Score uses two complementary lenses to assess consistency: magnitude stability (via coefficient of variation) and temporal consistency (via win rate decay over time).

The three quality dimensions — sample credibility, temporal distribution, and return consistency — are averaged with equal weight into a single reliability score. Each dimension is independent and tests a different failure mode: too few samples, regime clustering, and return instability. All three must be strong for the quality score to be high.

Sniper Score is the core output that drives ranking. It combines edge and quality into a single 0–100 conviction score, but the formula deliberately applies an asymmetric penalty: tickers with zero or negative edge are not simply ranked lower — they are mathematically suppressed regardless of quality. A perfectly reliable signal with no historical profitability is not a trading opportunity.

When edge is negative, the score becomes a product rather than an average — both factors multiply each other, creating a double-penalty that prevents quality alone from rescuing an unprofitable signal. When edge is positive, edge and quality are weighted equally, rewarding tickers that are both profitable and consistent. Scores below 50 are considered low conviction regardless of other metrics.

Every ticker in the Sniper table receives a verdict label derived directly from its Sniper Score. The verdict is displayed prominently on the per-ticker detail page alongside the Sniper Score and characteristic type.

For Muslim investors, each ticker in the Sniper table is tagged with its sharia compliance status sourced directly from Zoya Finance. The data is scraped on each run and displayed transparently in both the main table and the per-ticker detail page. Possible values are Halal, Not Halal, or Not Covered (no Zoya data available for that ticker).

Each ticker in the Sniper table is clickable and opens a dedicated detail page. The page consolidates all computed metrics into a structured breakdown to support manual review before any position decision.

The final output is a ranked table of all candidates that passed the Radar distance gate, sorted by SNIPER_SCORE from highest to lowest. Each row represents a ticker that has been fully processed through the edge and quality pipeline. Clicking any ticker opens its dedicated detail page with the complete per-ticker breakdown.

Planner reads from the same pipeline output as Sniper. All tickers that passed the Radar distance gate and completed the full backtest pipeline are displayed here, sorted by Sniper Score descending — identical ranking order to the Sniper table.

Each candidate is presented as a card with a checkbox on the left. The card surface shows everything needed for a quick first-pass judgment without opening any popup. The checkbox color reflects selection state: orange when unselected, green when selected.

EV represents the probability-weighted average outcome of a single trade at the selected time horizon. It accounts for both the frequency of wins and the magnitude of each outcome — a more complete picture than win rate or average return alone.

In practice, EV computed this way is mathematically equivalent to AVG_ret — the simple average return across all signal days at that horizon. AVG_ret is displayed on the card as a directly interpretable number: the average percentage return historically generated by this signal at this hold period. The full EV formula is shown in the Stress Test popup for transparency.

When one or more tickers are selected via checkbox, each selected ticker receives a portfolio allocation proportional to its Sniper Score relative to the total score of the selected basket. Higher conviction tickers automatically receive a larger slice. Allocation recalculates in real-time as tickers are added or removed.

Unselected tickers show a dash in the allocation field. The sticky footer displays the full basket summary: all selected tickers with their allocation chips, a count of selected positions, and the portfolio-level average EV.

Clicking the EV/Alloc box on any card opens a per-ticker stress test popup. This is the second layer of review — it breaks the EV figure down into its underlying components across all three time horizons, so the user can see not just what the expected return is, but how it is constructed and whether it holds across different hold periods.

Clicking the ticker name on any card navigates to the full Sniper detail page for that ticker (documented in 3.12). This is the deepest level of review available — the complete backtest history, regime check, quality breakdown, and conviction verdict. Intended for cases where the card and stress test alone are not sufficient to make a confident entry decision.

The Planner's decision output is not a single table but a live basket built interactively by the user. The three main outputs — EV per ticker, allocation per ticker, and portfolio average EV — together form the execution blueprint for the selected positions.

Data is organized by calendar month, starting from January 2026. Each month has its own view accessible from the landing page. Rows are trading days only — weekends and market holidays are excluded. Columns are tickers, each split into three sub-columns. One new row is appended every trading day automatically.

A ticker enters the Tracker the first day its DISTANCE_% exceeds zero — meaning it passed the Radar gate and became a Sniper candidate. From that point, a 20-trading-day window is opened. The ticker remains tracked for the full 20 days regardless of whether it continues to clear the Radar threshold on subsequent days.

If the ticker re-triggers (DISTANCE_% > 0 again) before the window closes, the window resets from that new date. A ticker exits the Tracker only when 20 trading days have elapsed since its last positive distance reading. This ensures the full forward return period referenced by the Sniper backtest is always observable.

Each tracked ticker exposes three daily data points in the matrix. Together they provide a complete picture of signal strength, price movement, and conviction status for every day the ticker is in the tracking window.

Several tools are provided to make the matrix easier to navigate and analyze without leaving the page.

A legend is displayed below the matrix for quick reference.

The simulation operates under a fixed, unchanging ruleset. No overrides, no manual selections, no characteristic or sharia filtering — the system picks mechanically based on rank alone.

Each 20-day hold cycle is called a batch. Batches are numbered sequentially from Batch 1. The system runs one batch at a time — there is no overlap between batches and no idle cash between cycles. The equity is always fully deployed.

Performance is benchmarked against SPY — the SPDR S&P 500 ETF, which tracks the S&P 500 index and is the most widely traded US equity benchmark. SPY is chosen because it is directly tradable (unlike the index itself), making the comparison an apples-to-apples evaluation against a real passive alternative.

SPY is rebased to 100 at the same start date (January 2, 2026) and rebalanced on the same cycle dates as the Alpha Executor — though SPY rebalances have no practical effect since the holding is always SPY. This ensures both equity curves share an identical starting point and time scale.

The top section of the dashboard displays five headline figures updated daily after market close.

The equity curve plots the full daily history of both the portfolio and SPY from inception to today. It is the primary visual for assessing strategy behavior over time.

A set of risk and return statistics computed across the full daily return history of the simulation. Updated daily after close.

Displays the five tickers currently held in the active batch, with their key position details. The table is static within a batch — holdings do not change until the next rebalance day.

A summary of the three most recently completed batches, displayed for quick historical context without needing to inspect the full equity curve.