Skip to content

Ukd1796/Financial_lab

BranchesTags

Repository files navigation

Financial Lab

NSE Indian Equity Backtesting & Strategy Research System

Research lab for modular, regime-aware trading strategy design on Indian equity markets. All results are net of 0.10% commission + 0.05% slippage per side. Paper trading only.

Current Results (March 2026)

Universe: 150 symbols (Nifty 50 + Nifty Next 50 + Nifty Midcap 50) Pipeline: DynamicUniverseAgent (150 → top 80) → per-strategy filter (80 → top 20)

Sharpe / Return% by Period

Strategy Full 2018–24 Bull 2019–20 Crash 2020 Recovery 2020–21 Bear 2022 Recent 2022–24
DualMA SMA20/50 1.29 / +216% 0.40 / +4% 1.38 / +24% 2.55 / +118% 0.56 / +8% 1.73 / +81%
Breakout 10d 1.12 / +155% 0.49 / +6% 1.49 / +25% 2.48 / +102% 0.39 / +4% 1.15 / +41%
QuietBrk 20d 1.05 / +146% 0.98 / +14% 2.00 / +35% 2.80 / +129% -0.67 / -10% 0.67 / +22%
TrendPB 3% 0.63 / +42% 0.39 / +3% 1.49 / +18% 1.27 / +29% -0.90 / -9% 0.44 / +8%
TrendPB 5% 0.80 / +39% 0.97 / +5% 1.80 / +16% 1.23 / +22% -0.34 / -2% 0.89 / +12%
RSI-MR os=10 -0.13 / -13% -0.34 / -4% 0.78 / +10% 1.06 / +25% -0.95 / -12% -0.37 / -11%
RSI-MR os=5 0.25 / +15% 0.21 / +2% 0.86 / +11% 1.37 / +33% -0.66 / -9% -0.11 / -5%

Good = Sharpe > 1.0 | OK = Sharpe 0.3–1.0 | Bad = Sharpe < 0.3

Key Observations

Regime-strategy mapping is now clear:

Regime Best Strategy Sharpe Why
Crash + Recovery QuietBrk 20d 2.00–2.80 High-vol upward moves clear 20d highs cleanly
Recovery + Recent DualMA 1.73–2.55 Sustained uptrend; golden crosses stay intact
Low-vol Bull QuietBrk 20d 0.98 0.8% threshold catches moves Breakout 10d misses
Bear DualMA only 0.56 Exits on cross-under; all other strategies lose
Mixed/Recent Breakout + DualMA 1.15–1.73 Both handle mixed volatility well

QuietBrk 20d is the biggest new finding: it dominates Crash and Recovery (Sharpe 2.00 and 2.80) and is competitive in slow-bull (0.98) but suffers in Bear (-0.67). It should be paired with a regime gate or adaptive weight — not run unconditionally.

DualMA is the most consistent bear-survival strategy (+8% in Bear, +81% in Recent, +118% in Recovery). Its golden-cross filter naturally avoids entries when the market is broadly declining.

RSI-MR remains structurally weak outside Recovery. The sma_cross_age >= 10 filter (R2) halved bear losses but RSI-MR 10 is still negative full-period. RSI-MR 5 (+15% full period) is the better variant if one must run it.

System Architecture

Market Data (SQLite)
       │
       ▼
DynamicUniverseAgent          ← bulk preloads all 150 symbols, scores by activity
  (150 → top 80 candidates)     signals: relative_volume, daily_return, sma_cross_age,
                                  sma_20_above_sma_50, return_3d, rolling_vol_5d
       │
       ▼
Per-Strategy Universe Filter  ← strategy-specific re-ranking of the 80 candidates
  (80 → top 20 symbols)         BreakoutFilter     : vol spike + large price move
                                 PullbackFilter     : uptrend + quiet pullback
                                 MeanReversionFilter: uptrend + oversold (sma_cross_age≥10)
                                 DualMAFilter       : fresh golden cross (sma_cross_age 1–5)
       │
       ▼
MarketObserverAgent           ← computes per-symbol indicators from OHLC history
  (lazy preload per symbol)     sma_5/10/20/50, atr_5/14, rsi_2/3,
                                 high_10d, high_20d, regime classification
       │
       ▼
Strategy.decide()             ← pure rule-based signal generation
  BreakoutMomentumStrategy      entry: price > high_10d  |  exit: price < sma_10
  QuietBreakoutStrategy         entry: price > high_20d  |  exit: price < sma_20
  TrendPullbackStrategy         entry: pullback in uptrend  |  exit: price > sma_20×1.05
  DualMovingAverageStrategy     entry: sma_20 crosses above sma_50  |  exit: cross below
  RSIMeanReversionStrategy      entry: RSI_3 < threshold  |  exit: RSI overbought or time-stop
       │
       ▼
RiskAgent                     ← position sizing + regime gate + breadth circuit breaker
  vol-adjusted sizing           risk_per_trade = 0.5% of portfolio / (2 × ATR)
  ATR stop (2×ATR)              hard stop checked on every HOLD decision
  regime filter                 e.g. UPTREND_ONLY blocks entry in DOWNTREND stocks
  breadth circuit breaker       blocks BUY when >40% of universe in DOWNTREND (R1)
       │
       ▼
ExecutionAgent + Portfolio    ← fills at close price with commission + slippage
       │
       ▼
EvaluationAgent               ← Sharpe, MaxDD, profit factor, win rate, trade count

Strategy Files

File Strategy Entry Signal Exit Signal Horizon
app/strategy/breakout_momentum.py BreakoutMomentumStrategy price > high_10d price < sma_10 3–10d
app/strategy/quiet_breakout.py QuietBreakoutStrategy price > high_20d price < sma_20 5–20d
app/strategy/trend_pullback.py TrendPullbackStrategy pullback in confirmed uptrend price > sma_20×1.05 5–15d
app/strategy/dual_ma.py DualMovingAverageStrategy sma_20 crosses above sma_50 sma_20 crosses below sma_50 weeks–months
app/strategy/rsi_mean_reversion.py RSIMeanReversionStrategy RSI_3 < threshold RSI overbought or max_hold_days 1–7d

Universe Filter Files

Class Strategy Key Filters
BreakoutUniverseFilter Breakout 10d rel_vol > 1.5, abs_return > 1.5%
BreakoutUniverseFilter (relaxed) QuietBrk 20d rel_vol > 1.2, abs_return > 0.8%
PullbackUniverseFilter TrendPB sma_20>sma_50, slope+, return_3d in [-12%, -1.5%], low volume
MeanReversionUniverseFilter RSI-MR sma_20>sma_50, sma_cross_age≥10, return_3d < -3%
DualMAUniverseFilter DualMA sma_20>sma_50, sma_cross_age 1–5, rel_vol > 0.8

Improvements Applied (R-series)

ID Change Impact
R1 Breadth circuit breaker 60% → 40% Breakout Bull improved -0.01 → +0.49; RSI-MR Bear halved losses
R2 sma_cross_age >= 10 in MeanReversionFilter RSI-MR full-period -29% → -13%; Bear -21% → -12%
R3 sma_cross_age >= 15 in PullbackFilter Reverted — harmed TrendPB Recovery without fixing Bear
B1 DualMA strategy with DualMAUniverseFilter +216% full period; best bear-survival strategy
B2 QuietBreakoutStrategy (20d, relaxed filter) Best Crash/Recovery performer (2.00/2.80 Sharpe)

Next Steps

Priority 1 — Adaptive Strategy Selector (LLM meta-layer)

The regime-strategy mapping is now clear enough to build a meta-layer that dynamically allocates capital between strategies based on current market conditions.

Design (see docs/adaptive_strategy_selector.md for full spec):

Weekly: RegimeSnapshot → AdaptiveStrategySelector (Claude API) → strategy weight vector
Daily:  weight vector → MultiStrategyRouter → RiskAgent (weight scales position size)

The LLM receives the current regime distribution (% uptrend/downtrend, ATR level, breadth) plus the historical Sharpe table above and outputs a capital weight per strategy. Called weekly (not daily) to avoid latency overhead.

Pre-requisites before building the selector:

  1. RSI-MR HOLD emission — RSI-MR does not emit HOLD for held positions, silently disabling the ATR stop. Fix: add explicit HOLD in the in_position branch (same pattern as BreakoutMomentumStrategy and TrendPullbackStrategy).

  2. QuietBrk 20d Bear regime gate — QuietBrk loses -10% in Bear. Either:

    • Add _UPTREND_ONLY as its allowed_regimes (currently _TREND_AND_SIDEWAYS), or
    • Let the adaptive selector weight it to 0 in bear conditions.

  3. MultiStrategyRouter — wraps all strategies, merges decisions (SELL overrides BUY, highest-weight strategy wins per symbol). Add weight and source fields to Decision.

Expected impact: In Recovery 2020-21, QuietBrk (2.80) and DualMA (2.55) both dominated. A selector routing 60% to QuietBrk + 30% to DualMA in that regime would significantly outperform running either alone at full capital. In Bear 2022, routing 80% to DualMA and 20% to Breakout (and 0% to everything else) avoids all the losing strategies.

Is the LLM the right tool? Yes, for this specific task:

  • 5 strategies × 4 regime dimensions = too many combinations to hand-tune
  • The mapping is pattern-matching with interacting variables, not a deterministic rule
  • Claude already "knows" the trade-offs (provided via the prompt's performance table)
  • A rules-based version requires 50+ manually tuned thresholds; the LLM infers them

See docs/adaptive_strategy_selector.md for full architecture, prompt design, MultiStrategyRouter pseudocode, and implementation order.

Priority 2 — TrendPB Structural Fix or Replacement

TrendPB 3% is still -9% in Bear 2022. The sma_cross_age >= 15 filter (R3) was reverted because it harmed Recovery without fixing Bear. Two options:

  • Fix: Add volume confirmation on entry — only buy the pullback if volume on pullback days is below 20-day average. Distribution (high volume pullback) vs healthy rotation (low volume pullback) is the key distinction.
  • Replace: Donchian channel breakout (20d high + volume confirmation) — similar to QuietBrk but with a price-channel rather than SMA exit.

Priority 3 — RSI-MR 10 vs RSI-MR 5

RSI-MR 10 is -13% full period; RSI-MR 5 is +15%. They share the same universe filter and risk settings — the only difference is the entry threshold. Consider retiring RSI-MR 10 from the active strategy pool. Running two variants of the same broken strategy adds noise without diversification.

Priority 4 — QuietBrk 20d Bear Regime Gate

QuietBrk 20d loses -10% in Bear 2022 vs +35% in Crash and +129% in Recovery. Its allowed_regimes is currently _TREND_AND_SIDEWAYS. Switching to _UPTREND_ONLY (block entries when per-stock regime is DOWNTREND) may reduce Bear losses at the cost of fewer Crash entries — worth testing since the ATR stop already limits per-trade losses.

Running Experiments

# Activate environment
source finance/bin/activate

# Run all strategies across all periods
python3 run_experiments.py

Results print per-period with Sharpe, return, max drawdown, profit factor, win rate, and trade count. Each period independently preloads and caches data — runs are fully reproducible.

Repository Layout

app/
  backtest/
    engine.py          # BacktestEngine — daily loop, universe filtering, P&L tracking
    observer.py        # MarketObserverAgent — per-symbol indicator precomputation
  data/
    models.py          # OHLCRecord, MarketState data models
    repository.py      # MarketDataRepository — SQLite OHLC access
  evaluation/
    agent.py           # EvaluationAgent — Sharpe, MaxDD, profit factor
  execution/
    agent.py           # ExecutionAgent — fill simulation with costs
  features/
    indicators.py      # SMA, ATR, RSI implementations
  portfolio/
    engine.py          # PortfolioEngine — position state, equity valuation
    models.py          # Portfolio dataclass
  risk/
    agent.py           # RiskAgent — sizing, ATR stop, regime gate, breadth CB
  strategy/
    base.py            # BaseStrategyAgent interface
    breakout_momentum.py   # Breakout 10d
    quiet_breakout.py      # Quiet Breakout 20d (B2)
    trend_pullback.py      # TrendPullback v2
    dual_ma.py             # DualMA golden cross (B1)
    rsi_mean_reversion.py  # RSI Mean Reversion
  universe/
    dynamic_agent.py   # DynamicUniverseAgent — bulk 150-symbol scanner
    filters.py         # Per-strategy universe filters (Breakout/Pullback/MeanRev/DualMA)
    models.py          # UniverseCandidate dataclass
    agent.py           # UniverseSelectionAgent (legacy fallback)
docs/
  improvement_roadmap.md         # Full before/after analysis of R1/R2/R3/B1/B2
  adaptive_strategy_selector.md  # LLM meta-layer design spec
  roadmap_next_improvements.md
  experiment_results_analysis.md
run_experiments.py     # Main experiment runner — all strategies × all periods
scripts/               # Data ingestion scripts

Disclaimer

Research and educational purposes only. No financial advice. No live trading.

About

A stateful financial decision engine exploring autonomous systems, capital allocation, and risk-aware strategy design.

Resources

Readme

License

MIT license
Activity

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages