Signal Systems Methodology

1. Data

• Source: daily OHLCV per NEPSE symbol — historical bars from the open samirwagle/Nepse-All-Scraper dataset (MIT licence), with the most recent sessions overlaid from our own NEPSE feed. No fabricated bars.
• Universe: every symbol with ≥ 250 valid sessions (316 symbols in the latest build). Symbols whose data ends more than 90 days before the newest bar are excluded as stale/delisted.
• Sample period: 2024-09-01 → 2026-10-28 (≈ 2.2 years) — the published backtest window, filled from the live data below. Individual symbols start later; signals are evaluated from each symbol's 61st bar.
• Cash dividends are not received in the simulation — a stated limitation that slightly understates high-dividend stocks.

2. Data sanitisation

The feed contains bars with impossible OHLC — e.g. NBL on 2013-05-09 printed open 532 against a 143–145 traded range (an unadjusted corporate-action open). Such bars corrupt ATR, stops, and the corporate-action detector, so:

1. Bars with non-positive OHLC, high < low, or malformed dates are dropped; dates de-duplicated.
2. If open or close falls outside the bar's own [low, high], it is clamped into the range. Every repair is counted and published (bars_repaired).

3. Corporate-action adjustment

NEPSE enforces a ±10% daily circuit, so an overnight gap beyond ±10% is a bonus/split/rights adjustment, not a real move. On the sanitised series, if open / prevClose is outside [0.90, 1.10], all prior bars are reverse-adjusted: prices × ratio, volume ÷ ratio, turnover unchanged. No new entries are taken on the adjustment bar itself. Adjustment count is published.

4. Indicators (all causal — bar i uses data up to i only)

• SMA(n): simple moving average of closes.
• RSI(14): Wilder's RSI.
• ATR(14): Wilder's smoothing of the true range.
• Bollinger (20, 2): lower = SMA20 − 2σ.

5. Entry rules

Entries are taken at the signal day's close. One position per symbol; max 10 concurrent positions. All systems require 20-day average turnover above Rs 10 lakh.

Momentum — Donchian-50 breakout + trend filter

• close > SMA50 and SMA20 > SMA50; close prints a 50-session highest close; 55 ≤ RSI(14) ≤ 80.

Trend Relay — SMA20/SMA50 golden cross + volume

• close < SMA50 on ≥ 20 of the last 30 sessions; SMA20 crosses above SMA50; volume > 1.5 × 20-day average.

Reversal — oversold bounce

• RSI(14) < 30 and close < lower Bollinger Band; bullish candle (close > open) closing in the top half of its range.

6. Exits, stops, targets

• Stop-loss (fixed at entry): entry − 2×ATR(14) (Momentum, Trend Relay), entry − 1.5×ATR(14) (Reversal). If a bar's low touches the stop, exit at min(stop, close) — gap-through exits at the close, no fantasy fills.
• Momentum trail: exit at the close when it breaks the 20-session lowest close.
• Trend Relay trail: exit on SMA20 crossing back below SMA50, or 3 consecutive closes under SMA50.
• Reversal: exit at the close when close ≥ SMA20 or RSI(14) crosses above 55.
• T1/T2 targets are reference levels only — they are NOT traded. The backtest exits solely via the rules above (targets_traded: false in the data).

7. Position sizing and costs

• Start Rs 1 Cr; each position a fixed Rs 10 lakh notional (qty = floor(10,00,000 / entryPrice)). Size does not compound with portfolio growth — deliberately conservative, disclosed.
• 0.5% round-trip cost (broker + SEBON + DP, simplified): half at entry, half at exit.

8. How each statistic is calculated

Let V0 = Rs 1 Cr, Vf = final portfolio value (cash + open positions at last close − exit costs), T = 2.15 years (the published backtest window, filled from the live data below).

The equity curve marks the full portfolio to market weekly (cash + open positions − exit costs). Alerts are positions still open at the dataset's last date. The ledger lists every closed trade with entry/exit prices, stop, targets, exit reason, Rs and % P&L, and holding days (most recent 500 in the download).

9. Known limitations

1. Corporate-action factors are inferred, not official.
2. Fixed Rs 10L notional understates what compounding would show.
3. Cash dividends are not received.
4. Entries at the signal day's close assume the close is tradable; only the 0.5% cost models friction.
5. One realised history — regimes change; nothing here predicts the future.

10. Live forward record — paper trades from real alerts

Each signal page also carries a Live forward record: a forward paper-trade ledger built from real alerts, not from the backtest.

• Every live Entry alert becomes a Rs 10 lakh paper trade at the signal close.
• Exits fire on real forward OHLC — Target 1, stop loss, a ratcheting trailing stop, or a 20-session time exit. No lookahead: only bars after the entry date are used.
• P&L is net of 0.5% round-trip charges.
• Seeded from the engine's verifiable alert history: every distinct BUY alert found in git-committed signal JSONs on origin/main (2026-09-30 commits onward), each traced to its real alert record (system, symbol, signal date, published entry/stop/targets). Forward tracking started 2026-08-03; SIKLES was excluded because no verifiable alert record exists for it — never invented.
• Recomputed daily after market close by tools/track-live-signals.py, which rewrites live_track_record in each signal JSON. The pages render these numbers client-side, so they stay fresh with no page rebuild.
• Honesty rule: win/failure rates appear only after trades close. With zero closed trades the section says so plainly — never 0%, never the backtest rate.
• Small-sample discipline (R5): n = decided trades only — open trades are shown separately and never enter n, the win rate, or profit factor. Win rate = wins ÷ n, shown with a Wilson 95% interval; profit factor = gross wins ÷ gross losses (net of 0.5% charges). Below n=30 the too-few-trades badge shows on the record itself: the sample is too small for the win rate or profit factor to prove an edge.