Swing Trader

Swing Trader

RSI(14)-based mean reversion on liquid large caps. Buys names that are statistically oversold (RSI<30), exits when they revert (RSI>70) or after 10 trading days regardless. Sizing is conviction-scaled — deeper oversold = larger position (within max_position_pct).

Required: _shared/intent-schema.md, _shared/position-sizing.md §"Conviction-scaled". See references/decision-rules.md.

Strategy summary

• Universe: 8 liquid large caps + 2 broad ETFs.
• Entry: RSI(14) closes below 30 — and the current RSI is still below 30 at tick time (no chasing past signals).
• Exit: RSI(14) closes above 70, OR position is 10 trading days old, whichever first.
• Sizing: conviction = (30 - current_rsi) / 30 clamped to [conviction_floor, conviction_ceiling]. Position notional = buying_power * max_position_pct * conviction.
• Cadence: daily (one tick per RTH session).
• No stops. Exits are RSI- and time-driven. A SOUL-level loss circuit breaker still applies.

6-phase protocol

Phase 1 — Pre-flight

Strategist-side: check that time_window is active. Read ~/.hermes/profiles/autotrader/state/swing-trader.json for current open positions opened by this strategist (so we can age them for exit).

Phase 2 — Compose context

market-microstructure with: "Are we in a regime where mean reversion in large caps is reliable, or is momentum dominant? One line."

risk-architecture with: "Anything in current swing-trader positions that warrants early exit?" (Pass position list.)

Both into notes. Microstructure response can downgrade conviction by 0.5x if response includes the literal token "momentum_dominant" — handled in phase 4.

Phase 3 — Signal gather

For each symbol in allowlist:

• get_equity_historicals(symbols=[symbol], start_time=<30 trading days ago>, interval=day) — fetch 30 daily bars (enough for RSI(14) with warmup).
• get_equity_quotes(symbols=[symbol]) — for intent_price.

Compute RSI(14) from the bars. Standard formula:

gains  = sum of positive close-to-close changes over 14 bars
losses = abs(sum of negative changes) over 14 bars
RS  = avg_gain / avg_loss
RSI = 100 - 100 / (1 + RS)

Phase 4 — Decide

Exits first (check each currently-open swing position):

for pos in open_positions:
    age_days = today - pos.opened_at
    current_rsi = rsi[pos.symbol]
    if current_rsi >= overbought or age_days >= max_days:
        emit sell intent for pos.qty (shares); reason: "rsi_revert" or "max_age"

Entries (for each symbol NOT already held):

for symbol in allowlist:
    if symbol in open_positions: skip
    if rsi[symbol] >= oversold: skip
    conviction = clamp((oversold - rsi[symbol]) / oversold, conviction_floor, conviction_ceiling)
    if microstructure_says_momentum_dominant: conviction *= 0.5
    target_notional = buying_power * max_position_pct/100 * conviction
    qty = floor(target_notional / quote.last_trade_price)
    if qty == 0: skip
    emit buy intent: shares=qty, order_type=limit, limit_price=quote.ask_price * 1.001

Concurrent-positions cap: if entries would exceed max_concurrent_positions, take the lowest-RSI candidates first.

Phase 5 — Hand off to executor

Standard envelope.

Phase 6 — Emit & persist

For any status: placed buy, append to state file:

{
  "open_positions": [
    {"symbol": "AAPL", "opened_at": "<ISO>", "qty": 1, "entry_rsi": 22.4, "entry_price": 184.50, "order_id": "..."}
  ]
}

For any status: placed sell, remove from state. Reviewed-only ticks do not touch state.

Interactive mode

• "Is NVDA a swing buy today?" — run phases 1-4 for just NVDA, emit tick_decision with one intent (review-only).
• "Exit all swing positions" — refuse. Exit logic is rule-based; user can edit hold.max_days to 0 to force liquidation on next tick.
• "What positions does swing have open?" — return state file contents as a table.

What this strategist will never do

• Buy a symbol off the allowlist.
• Hold past hold.max_days (force exit on the next tick).
• Average down. If a position drops and RSI re-prints below 30, this strategist sees the symbol is already held and skips re-entry.
• Use stop-loss orders. Exits are rule-driven, not price-triggered.