Nifty50 Equity Forecasting Benchmark

📄 Read the full thesis → THESIS.md — complete methodology, mathematics, worked examples, plots, results, and references, written for a machine-learning audience.

A disciplined, tier-based time-series forecasting benchmark over 25+ years of daily data for the 50 constituents of India's NSE Nifty 50 index — from trivial baselines to per-series classical models to a global gradient-boosted model — plus a live forward forecast of the next ~3 months per stock.

Methodology mirrors the sibling nav_forecast project: establish a floor with dumb baselines, then only believe a fancier model if it actually beats that floor. The recurring lesson in finance is that RandomWalkWithDrift on price levels is brutally hard to beat, and that forecasting returns (stationary) is the honest framing.

• Data: Kaggle kalyan197/nifty50-stocks1999-2026-daily-ohlcv-and-fundamentals — ~88 MB, ~287K daily rows, 49 stocks (of the Nifty 50), Jan 1999 → Jan 2026, sourced from Yahoo Finance, CC0.
• Targets: close (price level) and log_return (stationary).
• Horizons: 5 / 20 / 60 trading days (≈ 1 week / 1 month / 1 quarter).
• Validation: walk-forward CV, 6 folds, 60-day test windows, expanding train.
• Metrics: MAE, RMSE, sMAPE, MASE (scale-free; mean and median reported).
• Hardware: runs end-to-end on a Raspberry Pi 5 (4 cores, 8 GB) with thermal throttling.

🏁 Results at a glance

Mean MASE across 49 stocks × 6 walk-forward folds — lower is better; MASE < 1 beats the in-sample seasonal-naive. The winning model is shown per cell:

Target	h = 5 d	h = 20 d	h = 60 d
close — price level	RWD 1.92	LightGBM 3.39	AutoCES 5.52
log_return — stationary	HistoricAvg 0.40	AutoETS 0.39	AutoETS 0.39

Takeaway: at every horizon the best model only ties the trivial floor — RandomWalkWithDrift on levels, "predict-the-average-return" on returns — to within ~1–2%. Model complexity did not win. The full 10-model × 3-horizon tables for both targets are in the 🏆 Leaderboard section below.

---

The dataset

One combined long-format file plus a summary and metadata:

File	What it is
nifty50_historical_data.csv	Main panel — daily OHLCV + derived + fundamentals for all 49 tickers
nifty50_summary_statistics.csv	One-row-per-stock summary over the full period
metadata.json	Collection metadata

Main file columns: Date, Ticker (e.g. RELIANCE.NS), Company_Name, Sector, Open, High, Low, Close (split/dividend-adjusted), Volume, Dividend, Stock_Split, Daily_Return, Volatility_20D, MA_50, MA_200, Market_Cap, PE_Ratio, Forward_PE, PEG_Ratio, Price_to_Book, Dividend_Yield, EPS, Beta, 52Week_High, 52Week_Low.

Run ./run.sh eda.py for the full data-understanding report. Headline EDA findings are injected below after each run:

• Shape: 287,310 rows × 26 columns, 49 tickers (the dataset ships 49 of the Nifty 50), daily, 1999-01-01 → 2026-01-30.
• History per ticker: min 2,027 · median 5,858 · max 6,770 rows. All 49 reach ≥ 2026-01-01, so all 49 pass the eligibility filter.
• Sectors (13): Financials 10, IT 6, FMCG 5, Automobile 5, Metals 4, Pharma 4, Infrastructure 3, Cement 3, Energy 3, Consumer Durables 2, Power 2, Telecom 1, Healthcare 1.
• Leakage proof ✅: every one of the 10 fundamentals (Market_Cap, PE_Ratio, Forward_PE, PEG_Ratio, Price_to_Book, Dividend_Yield, EPS, Beta, 52Week_High, 52Week_Low) is constant within 49/49 tickers → confirmed point-in-time snapshots → dropped. (PEG_Ratio is 100% null anyway.)
• Daily_Return exactly equals Close.pct_change() (mean |diff| = 0.0) — it is just a convenience column; we recompute trailing features inside the models.
• Survivorship: 29/49 tickers list after 2000 (e.g. HDFCLIFE & SBILIFE 2017, LTIM 2016, COALINDIA 2010) — the panel is current constituents only.

⚠️ Two hazards this project handles explicitly

1. Fundamentals are point-in-time snapshots, not history. PE_Ratio, Market_Cap, EPS, Beta, Forward_PE, PEG_Ratio, Price_to_Book, Dividend_Yield, 52Week_High/Low are the current (2026) values repeated on every historical row. eda.py proves this (n_unique == 1 per ticker). Using them as time-varying features is lookahead leakage, so they are dropped from the modelling panel. Only Close (→ close, log_return) and Sector (static) are used.
2. Survivorship bias. The panel contains only the current index members, so long-dead names are missing and winners are over-represented. This is not fixable from the data and is reported as a limitation.

The pre-computed Daily_Return / Volatility_20D / MA_50 / MA_200 have unknown provenance, so trailing features are recomputed inside the models rather than trusted.

---

Methodology

See methods.md for the full tier roadmap. In short:

Tier	Models	Idea
0 — baselines	Naive, SeasonalNaive(5), RandomWalkWithDrift, HistoricAverage, WindowAverage(20)	The floor every later tier must beat.
1 — classical	AutoARIMA, AutoETS, AutoTheta, AutoCES (season 5)	Per-series Box-Jenkins / state-space / theta. Feasible at 50 series.
2 — global ML	LightGBM (mlforecast): lags + rolling stats + sector	One model pools all tickers; usually wins on returns.

Live forecast: models are refit on all history and projected ~60 business days forward — RandomWalkWithDrift + AutoETS (with 80/95% intervals) on levels, LightGBM on returns integrated to a price path.

---

🏆 Leaderboard

Walk-forward CV, cross-fold + cross-stock means; lower MASE is better (MASE < 1 beats the in-sample seasonal-naive). Regenerated by build_scoreboard.py (full interactive version in scoreboard.html).

Target: close

Horizon 5 trading days — ranked by MASE (lower is better)

Rank	Model	MASE	MASE (median)	MAE	RMSE	sMAPE
1	tier0:RWD	1.922	1.370	47.67	53.92	0.007875
2	tier1:CES	1.927	1.358	47.44	53.76	0.007889
3	tier1:AutoTheta	1.927	1.388	47.58	53.92	0.007895
4	tier0:Naive	1.933	1.363	48.07	54.43	0.00793
5	tier2:lgb	1.934	1.383	52.09	58.29	0.007918
6	tier1:AutoETS	1.951	1.376	48.96	55.38	0.007978
7	tier1:AutoARIMA	2.023	1.436	50.33	56.83	0.008207
8	tier0:SeasonalNaive	2.482	1.819	61.41	68.82	0.01021
9	tier0:WindowAverage	3.117	2.258	88.65	93.5	0.01291
10	tier0:HistoricAverage	86.628	86.932	2241	2242	0.5354

Horizon 20 trading days — ranked by MASE (lower is better)

Rank	Model	MASE	MASE (median)	MAE	RMSE	sMAPE
1	tier2:lgb	3.393	2.706	96.41	111	0.0139
2	tier1:CES	3.393	2.653	92.83	107.7	0.01398
3	tier0:RWD	3.416	2.747	93.74	108.6	0.01403
4	tier1:AutoTheta	3.431	2.765	93.74	108.7	0.0141
5	tier1:AutoETS	3.434	2.729	94.25	109.2	0.01418
6	tier0:Naive	3.464	2.750	94.53	109.5	0.01427
7	tier0:SeasonalNaive	3.831	2.947	100.5	117.1	0.01575
8	tier1:AutoARIMA	4.104	2.820	116.9	135.7	0.01627
9	tier0:WindowAverage	4.107	3.125	117.9	131.3	0.01706
10	tier0:HistoricAverage	87.391	88.489	2260	2261	0.5376

Horizon 60 trading days — ranked by MASE (lower is better)

Rank	Model	MASE	MASE (median)	MAE	RMSE	sMAPE
1	tier1:CES	5.521	4.580	160.5	188.4	0.02277
2	tier1:AutoETS	5.552	4.334	161	188.5	0.02305
3	tier0:RWD	5.579	4.314	162.3	189.8	0.02293
4	tier2:lgb	5.582	4.568	166.2	193.6	0.02284
5	tier1:AutoTheta	5.602	4.357	162.9	190.4	0.02302
6	tier0:Naive	5.713	4.411	165	192.8	0.02355
7	tier0:SeasonalNaive	6.010	4.873	169.3	198.5	0.02471
8	tier0:WindowAverage	6.097	4.927	179.3	206.6	0.0251
9	tier1:AutoARIMA	7.978	4.648	243	283.3	0.03057
10	tier0:HistoricAverage	88.531	87.603	2287	2290	0.5405

Target: log_return

Horizon 5 trading days — ranked by MASE (lower is better)

Rank	Model	MASE	MASE (median)	MAE	RMSE	sMAPE
1	tier0:HistoricAverage	0.396	0.342	0.008958	0.01095	0.8582
2	tier1:AutoETS	0.396	0.346	0.008962	0.01096	0.86
3	tier2:lgb	0.397	0.347	0.008996	0.01102	0.8355
4	tier1:AutoARIMA	0.400	0.345	0.009061	0.01105	0.8591
5	tier0:WindowAverage	0.408	0.350	0.009229	0.01129	0.797
6	tier1:AutoTheta	0.414	0.371	0.009386	0.01146	0.8048
7	tier1:CES	0.562	0.470	0.01282	0.01489	0.7405
8	tier0:SeasonalNaive	0.577	0.528	0.01306	0.01605	0.7328
9	tier0:Naive	0.580	0.486	0.01323	0.0153	0.7322
10	tier0:RWD	0.581	0.486	0.01323	0.0153	0.7323

Horizon 20 trading days — ranked by MASE (lower is better)

Rank	Model	MASE	MASE (median)	MAE	RMSE	sMAPE
1	tier1:AutoETS	0.388	0.374	0.008723	0.01142	0.8655
2	tier0:HistoricAverage	0.388	0.374	0.008725	0.01142	0.8634
3	tier1:AutoARIMA	0.391	0.375	0.008794	0.0115	0.8706
4	tier2:lgb	0.392	0.381	0.008808	0.01152	0.8328
5	tier0:WindowAverage	0.403	0.390	0.009066	0.01181	0.8065
6	tier1:AutoTheta	0.404	0.390	0.009086	0.01183	0.8041
7	tier1:CES	0.560	0.478	0.01278	0.01539	0.7429
8	tier0:SeasonalNaive	0.569	0.530	0.01283	0.0162	0.7324
9	tier0:Naive	0.575	0.487	0.01309	0.01568	0.7335
10	tier0:RWD	0.576	0.488	0.0131	0.01569	0.7337

Horizon 60 trading days — ranked by MASE (lower is better)

Rank	Model	MASE	MASE (median)	MAE	RMSE	sMAPE
1	tier1:AutoETS	0.391	0.380	0.008817	0.01191	0.865
2	tier0:HistoricAverage	0.392	0.382	0.008822	0.01191	0.8635
3	tier1:AutoARIMA	0.395	0.384	0.008888	0.01197	0.8739
4	tier2:lgb	0.398	0.390	0.008974	0.01205	0.819
5	tier0:WindowAverage	0.405	0.394	0.009116	0.01223	0.801
6	tier1:AutoTheta	0.405	0.392	0.00913	0.01225	0.7985
7	tier0:SeasonalNaive	0.566	0.531	0.01275	0.01632	0.7306
8	tier0:Naive	0.581	0.483	0.01329	0.0162	0.7382
9	tier0:RWD	0.583	0.482	0.01332	0.01624	0.7381
10	tier1:CES	0.689	0.472	0.01643	0.02184	0.75

⚙️ Does XGBoost / CatBoost / tuning help? (bake-off)

Short answer: no. Running every gradient-boosted tree head-to-head (3-fold walk-forward) plus an Optuna-tuned LightGBM (25 trials) — they cluster within ~2% and tie the random walk:

Target	h	LightGBM	XGBoost	CatBoost	LightGBM (tuned)	RWD floor
close	60d	5.63	5.65	5.58	5.70	5.58
log_return	60d	0.391	0.395	0.385	0.390	—

(MASE; full table in assets/bakeoff.csv.) The gaps are noise, and hyperparameter tuning did not help — Optuna's "best" config pushed min_data_in_leaf to ~500 (near-maximal regularisation), i.e. the search itself concluded "fit almost nothing." There is no XGBoost/CatBoost setting and no hyperparameter that extracts signal which isn't in the data. Reproduce: ./run.sh bakeoff.py.

🌊 Tier 3 — volatility: where a model finally beats naive

Returns are unpredictable in direction (≈50%), but their variance clusters — big moves follow big moves. So unlike price, volatility is forecastable. A GARCH(1,1) 1-step-ahead conditional-variance forecast beats a constant-volatility assumption on 98% of stocks (mean QLIKE −7.31 vs −6.84; lower is better). A simple trailing-20-day vol is about tied with GARCH (−7.46) — both capture the clustering; the point is that both crush "assume constant vol," whereas for returns nothing beats naive. This is the honest answer to "give me a model that isn't way off": forecast risk, not price.

GARCH (green) rises and falls with realized volatility (black) through every regime, while the constant baseline (red) is flat and blind. Reproduce: ./run.sh tier3_volatility.py.

---

🎯 Backtest: forecasts vs. actuals

There are two honest ways to compare forecasts to realized prices on the last 60 held-out trading days — they answer different questions and look completely different. Showing both is the whole point.

A. Multi-step (predict 60 days blind)

Train once, forecast the full 60 days with no feedback. Train ends at the dotted line; black is the realized price. Mean MAPE: RWD 4.45% · AutoETS 4.59% · LightGBM 4.91% — all near-flat, because that is the honest forecast for a near-random walk. Telling detail: the LightGBM here is trained with early-stopping on a validation set (the "assess how-far-off during training" step) and it stops at a single tree — the validation error can't improve beyond the drift, and directional accuracy is 50.8%. The model-assessment metric itself reports that there is no learnable signal in past prices. Where a forecast looks "off" (e.g. APOLLOHOSP fell ~12%) the stock moved — unpredictable from price history. All three use lagged features (LightGBM's lag1 is the #1 SHAP feature); flatness is the correct answer, not a failure to use recent data. Reproduce: ./run.sh backtest_plot.py.

B. One-step-ahead (rolling, re-anchored daily) — the forecast that tracks

At each day, predict tomorrow from the real data up to today, then re-anchor on the actual and step forward. Now the forecasts hug the actual — mean 1-step MAPE Naive 0.91% · AutoETS 0.94% · LightGBM 0.91% (vs ~4.5% multi-step). This looks like the forecast people expect — but it is not skill:

• Directional accuracy ≈ 50% (LightGBM 50.2%, AutoETS 50.9%) — a coin flip on the direction of the next move.
• LightGBM exactly ties Naive (0.91%) — the tuned tree adds nothing over "predict yesterday."

The lines track only because they predict ≈ yesterday's price (they're one day behind). That is the efficient-market result made visual: you can be ~99% accurate on the level and still have zero ability to call the move. Reproduce: ./run.sh backtest_onestep.py.

---

📈 Live forward forecasts

Next ~60 business days per stock — black = history, dashed blue = RandomWalkWithDrift, red = AutoETS (with 80% band), green = LightGBM (levels, differenced). Full gallery in assets/forecasts/.

Full 49-stock gallery in assets/forecasts/ and scoreboard.html.

---

Project structure

equity_forecast/
├── setup.sh                 one-shot: shared venv on SSD + pinned deps + Kaggle download
├── run.sh                   wrapper: activate venv + headless matplotlib
├── requirements.txt         pinned, ARM/py3.13-proven stack
├── data.py                  CSV → parquet panel (Nixtla long format) + static
├── splits.py                walk-forward CV + eligibility
├── metrics.py               multi-horizon MAE/RMSE/sMAPE/MASE
├── eda.py                   dataset understanding + leakage/survivorship proof
├── thermal.py               Pi thermal-throttle helper (no-op off-Pi)
├── tier0_baselines/run_baselines.py
├── tier1_classical/run_classical.py
├── tier2_global_ml/run_global.py
├── live_forecast.py         refit-on-all + forward charts
├── build_scoreboard.py      leaderboard → scoreboard.html + README
├── methods.md               tier roadmap
└── assets/forecasts/        committed forecast charts

Data, venv, and per-tier metric CSVs live on /mnt/ssd/equity_forecast/ (kept out of git); the committed artifacts are the code, scoreboard.html, and assets/.

How to run

./setup.sh                                   # venv + deps + dataset (one time)
./run.sh data.py materialise                 # build parquet panel
./run.sh eda.py                              # understand the data
./run.sh tier0_baselines/run_baselines.py    # baselines
./run.sh tier1_classical/run_classical.py    # classical
./run.sh tier2_global_ml/run_global.py        # global LightGBM
./run.sh live_forecast.py                    # forward forecast + charts
./run.sh build_scoreboard.py                 # leaderboard + scoreboard.html

Key findings

All tiers complete. One-line takeaway: on price levels nothing meaningfully beats RandomWalkWithDrift, and on returns the best a model does is recover the drift — where tier-1 AutoETS and the trivial HistoricAverage baseline tie for first. Complexity did not pay off here.

1. On price levels the whole field sits within a whisker of RandomWalkWithDrift. At h=60 the classical state-space models just edge it — AutoCES (MASE 5.521) and AutoETS (5.552) beat RWD (5.579) — while at h=5 RWD is #1 and at h=20 it ties LightGBM. The spread across the top six models is < 2%. Translation: nothing meaningfully beats "yesterday's price + drift" on levels; the classical wins are real but within noise.
2. On levels, AutoARIMA is the weakest classical model (MASE 7.98 at h=60 vs ≈ 5.5 for ETS/Theta/CES) — its weekly-seasonal (m=5) stepwise search overfits daily prices and costs ~25 min/fold. ETS/Theta/CES are both better and far cheaper.
3. On returns, "predict the drift" wins — and AutoETS does it best. At h=20/60 tier-1 AutoETS (MASE 0.388 / 0.391) leads, tying HistoricAverage and just ahead of LightGBM (0.392 / 0.398); all cluster at ≈ 0.39–0.41. Models that chase the last return (Naive / RWD / SeasonalNaive) sit at ≈ 0.58. The exception is AutoCES (≈ 0.56): it failed to fit some noisy return series and fell back to Naive (handled gracefully via fallback_model). Every return model beats the in-sample seasonal-naive (MASE < 1) — daily returns are otherwise near-random.
4. Global ML pooling helps marginally, not decisively. One LightGBM across all 49 tickers + sector matches the best baselines but does not clearly beat them — at 49 liquid large-caps with adjusted prices there is little extra point-forecast signal to extract at 1-week to 1-quarter horizons.
5. Sanity checks pass: HistoricAverage is a trap on levels (MASE ≈ 87 — the 25-year mean price is meaningless for a trending stock), and error grows monotonically with horizon (levels MASE 1.9 → 3.4 → 5.6 for 5 → 20 → 60d).

Limitations

• Survivorship bias — current constituents only.
• Fundamentals dropped — point-in-time snapshots would leak.
• No embargo/purging in the walk-forward CV yet (future tier-7 work).
• B-day grid + ffill approximates the NSE holiday calendar.
• Adjusted Close means dividends/splits are baked in (good for returns).

License & attribution

• Code: MIT — see LICENSE.
• Data: "Nifty50 Stocks (1999–2026) Daily OHLCV & Fundamentals" by Kaggle user kalyan197, released CC0 (public domain) and sourced from Yahoo Finance — kaggle.com/datasets/kalyan197/…. CC0 waives any attribution requirement, but the dataset author is credited here with thanks.

This is a research/educational benchmark, not investment advice.