ZENOJAS BiLSTM v1 — model card

What it does. ZENOJAS scores every stock in a ~940-name NSE universe each quarter with the probability that its price return over the next quarter will rank in the top 200 of the universe. The 25 highest-probability names form an equal-weight portfolio, rebalanced quarterly. Sector tilts from the Business-Cycle rulebook sit on top of this ranking; they are not applied in the published outputs because the current data set carries no sector classification.

Architecture. A learned ticker embedding is concatenated with 11 standardised features and passed through a single Bidirectional LSTM layer (16 units per direction, ReLU, dropout 0.2) into a sigmoid output. Loss is binary cross-entropy with Adam; 2.5 M parameters in total, of which the embedding is the bulk. The recurrent layer sees one time step per stock-quarter; the “16-quarter” tag refers to the lagged feature window used in preparation, not to a 16-step sequence.

Inputs (11 quarterly features per stock). Three-month price return (TMR), momentum (200-day minus 50-day moving average), beta and Jensen alpha versus the NIFTY Midcap 100, industry-relative price z-score, earnings yield (E/P), book-to-price (B/P), sales-to-price (S/P), a dividend-payer flag, debt-to-equity and EPS. These survived a VIF / correlation screen of a wider 30-ratio panel (ROCE, ROE, growth rates, EV/EBITDA, PEG and others were dropped for multicollinearity). Skewed features are power-transformed (Yeo-Johnson; Box-Cox for E/P and S/P) and standardised.

Data and training window. Quarterly fundamentals and prices from Bloomberg, Q4 FY05 onward. Training uses 2005 – 2021 (34,077 stock-quarters); the hold-out is a strict time-based split covering Mar-2022 to Dec-2024 (10,788 stock-quarters, 12 quarters). Two later quarters (Mar-2025, Jun-2025) were scored fully out-of-time and are shown separately as “live”.

Target. 1 if the stock’s next-quarter return rank is ≤ 200, else 0 (base rate ≈ 21 %).

Validation. Hold-out AUC 0.52; at a 0.5 threshold precision is 0.29 against a 0.21 base rate. Ranking the universe by score and buying the top 25 each quarter returned 40.9 % CAGR on the hold-out versus 26.6 % for the NIFTY Midcap 100 and 33.8 % for the equal-weighted universe, with a 21 % maximum drawdown and 58 % quarterly hit rate. Mean quarterly rank IC is close to zero and decile returns are not monotonic, so the outperformance should be treated as weak evidence rather than demonstrated skill.

Known limitations. Single hold-out window; no transaction costs, dividends or slippage; small-cap tilt inflates returns relative to a cap-weighted benchmark; features are Bloomberg-derived and a few stocks have missing inputs each quarter (imputed with the universe median); the ticker embedding cannot score names that were not in the training universe; no sector or liquidity constraints are applied. Outputs are research, not investment advice.