openenv.yaml

name: IndicatorsEnv
version: 4.0.0
description: >
  Multi-stock Relative Alpha MDP for NSE (Indian) equity analysis.
  At each step the agent observes 3 stocks from the same NSE sector and must
  identify which stock has the strongest relative momentum signal.
  It picks one stock (or passes with NONE) and declares a direction.
  Reward = (chosen stock period return - sector average) x direction x conviction x 50.
  Market-neutral design: random policy earns ~0 expected reward; a skilled policy
  earns consistently positive alpha by correctly identifying within-sector outperformers.
  Step spacing = GT window so there is zero ground-truth overlap between steps.

author: bawsi99
license: MIT

tags:
  - finance
  - stock-market
  - technical-analysis
  - reinforcement-learning
  - openenv
  - multi-stock
  - relative-alpha
  - market-neutral

tasks:
  - id: short_term_direction
    name: "Short-term Relative Alpha (Easy)"
    description: >
      5 steps, 1 trading day apart (= 1 week of daily observations).
      Observe 3 same-sector NSE stocks per step. Pick one or pass with NONE.
      Reward = (chosen 1-day return - sector 1-day avg) x direction x conviction x 50.
      GT threshold: +/-0.3% (1-day return window).
    difficulty: easy
    term: short
    episode_steps: 5
    step_spacing_days: 1

  - id: medium_term_direction
    name: "Medium-term Relative Alpha (Medium)"
    description: >
      10 steps, 5 trading days apart (= 10 weekly observations over 2.5 months).
      Signal history (RSI trends, price momentum) accumulates across steps
      to enable sequential multi-step inference.
      Reward = (chosen 5-day return - sector 5-day avg) x direction x conviction x 50.
      GT threshold: +/-1.5% (5-day return window).
    difficulty: medium
    term: medium
    episode_steps: 10
    step_spacing_days: 5

  - id: long_term_conviction
    name: "Long-term Risk-Constrained Alpha (Hard)"
    description: >
      15 steps, 20 trading days apart (= 15 monthly observations over 15 months).
      Spans multiple market regimes. Macro context (NIFTY50 trend, market regime)
      added to each observation. Episode terminates early if virtual capital
      drawdown exceeds 10%. Conviction calibration scored.
      GT threshold: +/-2.5% (20-day return window).
    difficulty: hard
    term: long
    episode_steps: 15
    step_spacing_days: 20

endpoints:
  reset: "POST /reset"
  step: "POST /step"
  state: "GET /state"
  tasks: "GET /tasks"
  grader: "POST /grader"
  baseline: "GET /baseline"
  health: "GET /health"
  websocket: "WS /ws"

action_space:
  type: object
  required:
    - stock
    - direction
    - conviction
  properties:
    stock:
      type: string
      description: >
        NSE symbol to trade (must be one of the 3 available_stocks in the observation),
        or "NONE" to skip this step.
      example: "HDFCBANK"
    direction:
      type: string
      enum: [Bullish, Bearish, NONE]
      description: >
        Bullish = bet chosen stock outperforms sector average (long alpha).
        Bearish = bet chosen stock underperforms sector average (short alpha).
        NONE = pass this step (reward = 0, excluded from grader scoring).
    conviction:
      type: number
      minimum: 0.0
      maximum: 1.0
      description: >
        Kelly fraction: proportion of virtual wealth wagered on this prediction.
        Higher conviction amplifies both gains and losses proportionally.
        Set to 0.0 when passing (direction = NONE).

observation_space:
  type: object
  properties:
    step:
      type: integer
      description: Current step number (1-indexed)
    max_steps:
      type: integer
      description: Total steps in this episode (5 / 10 / 15 by task)
    term:
      type: string
      enum: [SHORT, MEDIUM, LONG]
    sector:
      type: string
      description: "NSE sector shared by all 3 stocks (banking / it / pharma / fmcg / auto)"
    available_stocks:
      type: array
      items:
        type: string
      description: Exactly 3 NSE symbols to choose from this step
    stocks:
      type: object
      description: >
        Dict keyed by symbol. Each entry: current_price, rsi_14, rsi_trend,
        price_momentum_pct (cumulative return vs episode start), and a full
        indicators dict (moving_averages, macd, bollinger_bands, adx,
        stochastic, volatility, enhanced_volume, pivot_points).
    signal_history:
      type: array
      description: >
        Chronological log of past picks. Each entry: step, picked_stock,
        direction, ground_truth, correct, alpha_pct, reward, rsi_snapshot.
        Empty on step 1; grows by one entry each step.
    macro:
      type: object
      nullable: true
      description: >
        Task 3 only. Fields: nifty_trend (Bullish/Bearish/Neutral),
        nifty_return_20d (float %), market_regime (trending/ranging).

reward:
  formula: "(chosen_return - sector_avg_return) x direction_sign x conviction x 50"
  min: -1.5
  max: 1.5
  notes: >
    Market-neutral: sector average cancels broad market beta so a random agent
    earns ~0 expected reward. NONE action yields reward = 0 and is excluded
    from grader scoring. direction_sign is +1 for Bullish, -1 for Bearish.
    Conviction scales reward linearly (Kelly criterion semantics).