Backtesting

Backtesting

backtest.py answers one question with data: do BUY-scored stocks actually outperform SELL/HOLD-scored ones? It replays the scoring rules on history instead of trusting conventions.

Method

• Universe: ~100 NSE stocks (NIFTY 50 + Next 50 proxy)
• Window: 5 years of daily adjusted closes (~1,240 trading days)
• Rebalance: each month-end, score every stock with the exact technical rules from analyze_stock, using only data up to that date (no lookahead)
• Measure: forward 1/3/6/12-month returns per score bucket and per score quintile, benchmarked against NIFTY 50 mean forward returns over the same windows
• Buckets: BUY ≥60, SELL ≤20, else HOLD (same cutoffs as the server)

Current findings (beta weights)

1. BUY ≈ HOLD — the score's level doesn't separate winners at 6–12 months. Outperformance vs NIFTY is explained by survivorship bias, not skill.
2. Weak short-horizon momentum — top-quintile scores beat bottom-quintile at 1–3 months (e.g. +6.7% vs +4.4% 3m) — the only edge currently evidenced.
3. SELL never triggers — the technical-only score floor (~15) makes the ≤20 cutoff nearly unreachable; server SELL calls come from the fundamental half.
4. Non-monotonic long end — Q4, not Q5, had the best 6-month returns: long-horizon ranking power is absent.

Known limitations (stated in the output too)

• Technical rules only — yfinance has no point-in-time historical fundamentals
• Survivorship bias — universe is today's index members
• No transaction costs, slippage, or taxes

Outputs

• Console tables (buckets + quintiles + BUY−SELL spread)
• backtest_results.csv — every stock-month with score, bucket, forward returns
• backtest_summary.json — machine-readable summary

Next measurable steps

1. Per-rule ablation (score each rule in isolation; keep what separates quintiles)
2. Point-in-time fundamentals source to backtest the value half
3. Re-tune weights toward the evidenced short-horizon momentum signal