From `nse-options-mcp-research/signal_gate.py` — why a trading system should say "I don't know" most of the time, and the exact rule I coded. If your system always says BUY/SELL, ye padho.

The trap

Retail systems har bar signal dete hain. Confidence 55% pe bhi "BUY" bol dete hain. Result: V1 paper trading mei **31.6% win, −₹90.3k PnL**. The fix isn't a better model — it's a **gate** that refuses to vote when evidence is thin.

What a signal gate does

Input: market snapshot + session history. Output: ek decision:

Gate model se pehle aata hai. Model kuch bhi bole, gate veto kar sakta hai.

The ≥20 sessions rule

Guide ka hard rule: **no trade signal until ≥20 recorded sessions** in DuckDB history. Why? 1-2 sessions ka "pattern" noise hai. 20+ sessions se basic regime/volatility stats bante hain.

MIN_SESSIONS = 20
sessions = db.execute(
    "SELECT count(DISTINCT session_date) FROM market_raw").fetchone()[0]
if sessions < MIN_SESSIONS:
    return ("NO_TRADE", "INSUFFICIENT_HISTORY", f"{sessions}/{MIN_SESSIONS}")

Other gates

Now() IST handling

from datetime import datetime, timedelta
now_ist = (datetime.utcnow() + timedelta(hours=5, minutes=30)).replace(tzinfo=None)

DB naive IST store karta hai, isliye naive compare. Pehle aware/naive mismatch se tz error aata tha — fix kiya.

Real result

Mera system abhi **NO_TRADE / INSUFFICIENT_HISTORY** deta hai (sirf 1 session). Ye sahi hai. 5-min recorder 4 weeks mei 20+ sessions build karega, tab real probability niklegi.

Why this matters

Ek system jo NO_TRADE 90% bolta hai aur sahi 10% mei trades kare = survivable. Ek system jo har bar trade kare = margin call. Gate pehle, model baad mei.

Comparison

| Without gate | With gate |

|-|--|

| Always signals | Signals only when ready |

| 55% conf trades |

| −₹90k PnL | Protected capital |

| False confidence | Honest uncertainty |

Design principles

1. Gate model se independent hona chahiye (model bias na le).

2. Har reason logged hona chahiye (audit trail).

3. Kill-switch hard fail pe trip kare.

4. Session/holiday rules non-bypassable.

FAQ

**Q: 20 sessions kyun?** Basic stats ke liye minimum sample. Kam mei noise.

**Q: Gate model ko override kar sakta?** Haan, NO_TRADE har time model veto karega.

**Q: Live source down toh?** NO_TRADE, stale pe trade nahi.

**Q: Kill-switch manual?** Auto trip on hard failure + manual bhi.

Common mistakes

1. Gate ko model ke baad lagana — pehle lagao.

2. Session count ignore karna — 1 session pe trade karna.

3. Stale feed allow karna — 10 min old data pe trade.

4. Kill-switch na hona — hard fail pe bhi trade.

What I learned

V1 mei gate nahi tha, isliye 31.6% win pe bhi trade ho raha tha. V2 mei gate first line of defense hai. Ab system 4 weeks tak NO_TRADE bolega — wo loss nahi, protection hai.

*Research only. Not investment advice.*

Deep dive: gate logic in code

def evaluate(snapshot, db):
    # 1. Live source
    if not snapshot or snapshot.get("stale"):
        return ("NO_TRADE", "NO_LIVE_SOURCE", None)
    # 2. Session count
    sessions = db.execute(
        "SELECT count(DISTINCT session_date) FROM market_raw").fetchone()[0]
    if sessions < MIN_SESSIONS:
        return ("NO_TRADE", "INSUFFICIENT_HISTORY", f"{sessions}/{MIN_SESSIONS}")
    # 3. Holiday/weekend
    if is_nse_holiday(now_ist):
        return ("NO_TRADE", "HOLIDAY", None)
    # 4. Stale feed
    if (now_ist - snapshot["ts"]).seconds > 600:
        return ("NO_TRADE", "STALE_FEED", None)
    # 5. Model confidence
    conf = model.predict(snapshot)
    if conf < 0.60:
        return ("NO_TRADE", "LOW_CONFIDENCE", conf)
    return ("TRADE", "OK", conf)

Session build timeline

5-min recorder 2 symbols × ~124 rows = ~250 rows/session. 20 sessions ≈ 4 weeks market days. Tab tak system data collect karta hai, trade nahi.

Why 0.60 confidence threshold

V1 mei 55-64% confidence pe trade hue, sab loss. 0.60 conservative start hai. Jab data build hoga, calibration se threshold tune karna (isotonic).

Kill-switch design

def killswitch_tripped(reason):
    db.execute("INSERT INTO scan_log VALUES (?, 'KILLSWITCH', ?)",
               (now_ist, reason))
    return ("NO_TRADE", "KILLSWITCH", reason)

Hard failure (MCP crash, DB corrupt) pe trip. Manual reset only.

Real-world analogy

Gate ek bouncer hai club ke bahar. Model andar dance karna chahta hai, bouncer check karta hai: ID hai? (session) Source legit? (live) Time sahi? (holiday) Bouncer na bola toh andar nahi jaane deta.

My experience

Pehle bina gate ke model ko bharosa tha. 31.6% win dekh kar samjha ki problem model nahi, discipline thi. Gate ne discipline enforce kiya — ab system 4 weeks chup rahega, jo sahi hai.

*Research only. Not investment advice.*

Extended FAQ

**Q: Gate model se independent kyun?** Model apna bias defend karega. Gate neutral check kare.

**Q: 20 sessions bad mei kya?** Walk-forward validator real signal nikalta hai, tab TRADE allow.

**Q: Confidence threshold tune kaise?** Calibration set pe isotonic regression, false-positives minimize.

**Q: Manual override?** Sirf emergency kill-switch, normal NO_TRADE auto.

Common pitfalls (retail)

1. **Always-on signals** — 55% pe bhi trade. Capital wipe.

2. **No session gate** — 1 day data pe pattern seekh ke trade.

3. **Stale data trade** — 30 min purani snapshot pe decision.

4. **No kill-switch** — crash mei bhi orders.

Comparison: my V1 vs V2

| | V1 | V2 |

|--|-|-|

| Gate | None | 5-layer |

| Session min | 0 | 20 |

| Result | −₹90.3k | Protected |

| Confidence | 55% trades | <60% NO_TRADE |

What you should build

Agar tum apna system bana rahe ho:

1. Gate model se pehle lagao.

2. Session minimum set karo (20+).

3. Live source + stale check.

4. Kill-switch.

5. Har decision log karo.

Closing

Signal gate trading system ka seatbelt hai. Bina uske tum 31.6% win pe bhi trade karoge aur paise gawayoge. NO_TRADE bolna seekho — wo weakness nahi, strength hai.

*Research only. Not investment advice. SEBI compliance separate topic.*

Worked scenario

Monday 09:35 IST. Snapshot aaya. Gate evaluate:

1. Live source? YES (MCP up).

2. Sessions? 1/20 → **NO_TRADE / INSUFFICIENT_HISTORY**. Stop. Model call hi nahi hua.

Week 4, 21 sessions. Gate:

1. Live source? YES.

2. Sessions? 21/20 → pass.

3. Holiday? No.

4. Stale? 3 min old → pass.

5. Confidence? 0.67 → **TRADE**.

Ye difference hai gate wale aur bina gate wale mei.

Why I coded it this way

V1 ke sabse dardnaak loss tab aaye jab 1-2 sessions ka data tha aur model confident tha. Gate ne wo band kar diya. Ab system 4 weeks discipline se data collect karega, fir trade karega.

Recommendations

1. Gate ko model se alag rakho (separate module).

2. Har reason log karo — audit trail future debug ke liye.

3. Threshold conservative rakho pehle (0.60), bad mei tune.

4. Kill-switch non-bypassable ho.

*Research only. Not investment advice.*

The philosophy

Trading mei "not trading" bhi ek position hai. Zyada traders isko bhool jate hain. Gate system ko force karta hai ki wo apni limits jane. 90% time NO_TRADE bolna = 90% time capital safe.

Summary

Signal gate = system ka seatbelt. 5 layers (live source, sessions≥20, holiday, stale, confidence≥0.60) + kill-switch. V1 bina gate ke −₹90k gaya, V2 gate ke saath protected. Build it first, model later.

*Research only. Not investment advice. SEBI compliance separate topic.*

Action plan

1. Model se pehle gate likho.

2. Session minimum 20 rakho.

3. Live source + stale check lagao.

4. Kill-switch banao.

5. Har decision log karo.

Bina gate ke mat trade karo — V1 ne sikha diya ye.

Agar tumhare system mei bhi har bar signal aa raha hai, gate nahi hai. 5 layers lagao, dekho kitne signals survive karte hain. Shayad 90% NO_TRADE ho — wo hi sahi hai.

Final note

Gate design mei sabse zaroori: model ko veto power dena. Model chahe jitna confident ho, gate NO_TRADE bol sakta hai. Ye hierarchy retail systems mei missing hoti hai, aur isliye woh blow up hote hain.

*Research only. Not investment advice.*

Why I'm sharing this

Zyada "AI trading" content fake 90% accuracy bechta hai. Mera system abhi NO_TRADE de raha hai by design — wo honest hai. Gate lagana hi wo cheez hai jo retail ko pro bana ti hai. Build it, respect it.

*Research only. Not investment advice. SEBI compliance separate topic.*

Next: hum isi gate + data pe walk-forward validator chalayenge jab 20 sessions build honge. Tab real signals milengi.

Build the gate first. Your capital will thank you.

More From Shakti Tiwari