Zen: investment research infrastructure for Indian equities

I built Zen as research infrastructure for my own Indian equity process. It brings point-in-time market data, financial results and company filings into one archive, then uses that archive for two jobs: a weekday market brief and a research harness for testing stock-selection rules without hindsight.

I designed the investment process, data requirements, rules and validation framework and used AI-assisted coding tools to implement it. My focus was deciding what the system should test, what information was available at each point in time, how the results should be challenged, and how the output should support investment research. I was not trying to demonstrate software engineering.

The first question I wanted to answer properly was simple: if I screen Indian stocks on fundamentals and market factors, does the process actually add value beyond the market and beyond simply owning the same eligible universe? Most retail backtests cannot answer that honestly because they test against the companies that survive today. So the first job was building an archive that does not cheat, and a test harness that tries to break its own results.

The answer, in a backtest over the whole 7.6 years from Feb 2019: 30.7% a year against 14.4% for the Nifty 500 including dividends, and 22.0% for every stock that passed its filters, bought in equal amounts. It beat 96% of 500 random portfolios picked from that same list under the same rules, whose median made 20.7% a year, and its worst fall was -33.6% against -38.1% for the Nifty 500 and -47.9% for that list. On the last 3.6 years, which the design had never seen, it made 34.3% a year against 13.5% for the Nifty 500 and 21.0% for that list equally weighted; the run on record from 22 Sep 2026, measured from the close of 15 Feb 2023 rather than the open, and before two data faults were fixed with no stock-picking rule changed, gave 29.4% (below).

The rules were public before the test. Most of the return is the market and two well-known factors, and what is left after them is not yet statistically significant. What that does and does not prove is below.

It runs unattended on GitHub Actions and emails me two briefs. It places no orders. It produces evidence, and I make the decisions.

prices        5,360,616 rows · 2,895 sessions · 4,066 ticker symbols · Jan 2015 to Sep 2026
filings         788,772 announcements · 2,555 symbols · Jan 2022 to Sep 2026
fundamentals    105,919 quarterly results filings · 2,643 symbols · quarters ending Mar 2018 to Jun 2026
indices              28 NSE index series · Jan 2015 to Sep 2026

Counted up to 23 Sep 2026.

Most of the backtest and archive figures in this README are written by jobs/readme_numbers.py from the committed results and the database, and tests/test_readme_numbers.py fails if the README drifts from them. The few written by hand name the file or clarification they come from.

Where to look first

If you want	Go to
The live paper record (Zen v1, from 5 Oct 2026)	zen-research.pages.dev, a read-only dashboard rebuilt every weekday evening
What the emails look like	sample brief
The result, and what it does not prove	The result below
The headline and acceptance figures, each with the file and key it comes from	data/backtest/v1_final/baseline_report.md and its .json; the rest are in stats.json, verify.json, libcheck.json and nav.csv beside it
The rules, fixed before the test	research/strategy/v1-spec.md
Every hypothesis I've formally tested, including the failures	state/trials.jsonl
The look-ahead detector	zen/validation/leak.py
What I got wrong	Mistakes below

How I use it

Morning brief. I wanted the data to lead the research rather than the headlines. The brief reads the market first, identifies breadth, unusual volume, 52-week extremes and other moves, then promotes the filings and stories most relevant to what actually happened and to the themes I follow.

Strategy research. I wanted to know whether a process actually added value, so I tried to falsify it rather than optimise a backtest until it looked good. The rules are fixed before held-back testing, the universe is point-in-time, and the result is checked against the same eligible universe, random portfolios and factor attribution.

Paper portfolios. A separate private project keeps prospective paper records for Zen v1, a newer fundamental-rules strategy and my discretionary ideas. Decisions are recorded before execution and the historical backtest is kept separate from the live paper record. I keep that repository private because it contains my ongoing research and decision log; a read-only copy of its dashboard, with Zen v1's live record since 5 Oct 2026 beside its backtest, is public at zen-research.pages.dev.

---

The morning brief

Open a real one · sent every weekday at 07:00 IST

It isn't a news digest. It reads the archive first, works out what the session actually did, then goes looking for stories that explain it.

The data comes first, not the headlines. Breadth, the gap between the heavyweights and the median stock, 52-week extremes, size rotation, and every stock that traded far above its own 60-day volume. Whatever the archive flags is what decides which stories get promoted. A filing that explains an unusual move beats a better-sourced story about something that didn't happen.

It's ranked for one reader. Trust, corroboration and recency will give you a decent front page, but two people with opposite portfolios would get the same email. So stories are also scored against my own themes: AI and data centres, nuclear, solar, water, infrastructure, defence, import substitution. Each one carries a badge saying which theme matched. Theme affinity gets added to the score rather than multiplied, because a brief that only shows me what I already believe is worse than a generic one. I still need to know the market fell.

Keyword matching is the weak part, so it's bounded. Each theme separates terms that count on their own from terms that need backing up. One "data centre" is enough. One "power" isn't. Where the news API supplies entity tags the guessing stops, and the story carries real NSE tickers with sentiment.

The charts are HTML, not images. Gmail, Outlook and Apple Mail all block images by default for a sender you haven't whitelisted, so a brief built on attached PNGs is a wall of text for a lot of opens. The advance/decline bar, the 30-session breadth chart, the divergence gauges and the rotation table are all built from table cells and background colours. They render everywhere, with images off. Every bar carries its number as well, because colour on its own is unreadable for about one man in twelve.

Nothing breaks the send. No model key means extracted facts instead of prose. A dead feed means fewer stories. A broken chart means no chart. The email goes out.

The second email

A calibration run

Silent by default. It only sends when something fires, because an alert that arrives every day stops being an alert.

Strategies carry a validated flag. It defaults to false and only becomes true after a walk-forward backtest that survives the look-ahead detector and reports its trial count. While anything unvalidated is in the mix, the subject line says CALIBRATION, the sections read "ranked highest by the test screen" instead of "to buy", and a banner sits at the top telling me not to trade on it.

That's there because an earlier version headed a calibration run "To buy (15)" and I read it exactly as it was written.

---

Why survivorship bias is the whole problem

Most retail backtests in India run against whatever tickers are listed today. Companies that delisted, collapsed or got absorbed are just missing from the sample. So the strategy only ever gets measured against the survivors, and every result comes out better than it was.

It isn't a small effect. Counting ordinary equity only (ISIN security type 01, so no ETFs or rights entitlements), this archive holds 3,325 ticker symbols. Linked across renames and ISIN changes by zen/universe/identity.py, they belong to 3,050 companies, and 2,560 of those traded in NSE's main equity series in Sep 2026, up to the 23rd. That leaves 490 companies that were tradeable at some point since 1 Jan 2015 and are no longer quoted there. Of those, 35 still trade in NSE's trade-for-trade segment; the other 455 traded in no NSE series at all that month.

The universe on each date is whatever actually traded that day:

Month	Ordinary equity symbols trading
Mar 2015	1,476
Jun 2023	1,875
Sep 2026, to the 23rd	2,563

It's built from NSE daily bhavcopy files, which record every security that traded that day. The universe on any past date is what actually existed then, not what made it to now.

---

Decisions I'd defend

The database is derived, not stored. Git tracks one Parquet file a month and DuckDB gets rebuilt from them in seconds. Committing a growing binary every day would have bloated the repo inside a month. Monthly partitions instead of daily cut the file count from about 2,500 to 120 per decade, and they compress much better.

Signals have to argue against themselves. The Signal class won't construct without a stated counter-argument. If I can't name how a position fails, I haven't thought it through, so the constructor enforces that instead of relying on my discipline.

Rules decide, models explain. Story selection, deduplication and ranking are deterministic and I can inspect all of it. The language model only rewrites the chosen stories in plain English. If it's unavailable the brief still goes out.

Claims stay inside their evidence. The brief says "volume spikes our news sources do not account for". It does not say "no news exists". The feeds carry market news, not company filings, and the wording says exactly what was checked.

---

How I test it

The main test is hard to fool:

Run the strategy as of date T against the full archive. Run it again against an archive physically truncated at T. Both answers have to be identical.

Any look-ahead changes the answer. A mis-shifted column, a normalisation computed over the whole sample, a join that quietly pulls tomorrow's row. They all fail the same test and I don't have to anticipate them one by one.

tests/test_leak.py holds a strategy that ranks on forward returns, and it has to fail the detector. A detector that never fires proves nothing, so the suite asserts the failure. If that test ever passes, the harness is broken and everything downstream of it is worthless.

This exists because an earlier version of this work reported a 30% CAGR that turned out to be a one-day look-ahead bug. I built the harness before the strategy this time.

---

The result

A fundamentals-led screen, committed to this repo before it was tested, beat the market on data it had never seen, and over the whole period beat random picks from the same list. Most of its return is the market and two well-known factors, and what is left after them is not yet statistically significant.

The rules were written into research/strategy/v1-spec.md and committed on 21 September 2026. Everything from 15 February 2023 onward was locked in code, and on 22 September the configuration chosen in advance was run on it. The first attempt that day had a bug: it held the November 2022 portfolio frozen for four years instead of rebalancing. That attempt was discarded, the bug fixed and the test run again the same day with no rule changed (commit 6223419; v1 disclosure 40). That run stays on record as the pre-registered result.

Two faults in the archive came to light after it, both in the data rather than the rules. The code listing NSE's filings stopped at the first page that failed to load, so the archive held the March 2025 quarter for only 1,516 companies. It now holds 2,166, and the same pass fetched about 2,200 later documents, almost all revisions the regular updater had never collected (Clarification 36); the run was repeated on 23 September. And some filings tag their income statement only for a half-year or a year, which the parser stored as the quarter (Clarification 38, 24 September). A figure now counts as a quarter only if its period is at most 100 days long, and Clarification 39 settles how a quarter with no such figure is read, a case the two engines had read differently; the strategy's holdings are the same under both readings. The whole run, the held-back years included, was repeated on 25 and 26 September 2026 with no rule for choosing or weighting stocks changed. One exit detail was completed at the same time: a session on which a held stock trades only in NSE's trade-for-trade segment now counts as a trading day for the 20-session exit, as the 22 September fixes had intended but never delivered to the engine (v1 disclosure 34; v2-spec, Clarification to A2).

Run	What changed	Held back, a year	Same universe, equal weight, held back	Whole period, a year
22 Sep 2026	the one-shot test, as run	29.43%*	20.37%*	28.35%
23 Sep 2026	March 2025 quarter refilled, revised filings added (Clarification 36)	32.97%	21.03%	30.06%
25-26 Sep 2026	only quarter-length figures read as a quarter (Clarifications 38 and 39); the current result	34.31%	20.96%	30.68%

* From the close of 15 Feb 2023, the clock used at the time. The later runs start at the open, as Clarification 16 requires (Clarification 37).

At the 17 Nov 2025 rebalance the one-shot run bought five stocks, and three of them, CESC, FMGOETZE and IMPAL, were bought while the archive held their April-to-September figures as the September quarter; the 23 Sep run still bought IMPAL that way. The current run holds ASHOKLEY in place of IMPAL for that quarter and picks the same stocks as the 23 Sep run at every other rebalance. Counting trade-for-trade sessions as trading days changed five holdings of the equal-weight universe and none of the strategy's (trade_for_trade.csv).

The current run, on the repaired archive:

Open of 15 Feb 2019 to open of 18 Sep 2026, 7.6 years, of which the last 3.6, from the open of 15 Feb 2023, were held back:

	Whole period, a year	Worst fall	Held back, a year	Worst fall
The strategy	30.68%	-33.6%	34.31%	-23.8%
The same universe, equally weighted	22.03%	-47.9%	20.96%	-31.3%
Nifty 500	14.36%	-38.1%	13.50%	-18.6%
Nifty Midcap 150	20.38%	-38.5%	21.42%	-20.9%
Nifty Smallcap 250	20.14%	-45.5%	21.88%	-26.0%
Nifty200 Momentum 30 †	15.62%	-34.0%	15.92%	-31.6%
Nifty500 Value 50 †	22.97%	-53.4%	27.30%	-22.2%
Nifty200 Quality 30	11.91%	-29.1%	11.34%	-22.4%
Nifty100 Low Volatility 30	13.31%	-30.6%	12.75%	-18.1%
Nifty Alpha 50	23.48%	-38.3%	24.45%	-31.3%

Indices are NSE's total return indices, dividends included, on the strategy's clock; where NSE printed no open for a factor index, its parent index's overnight move stands in (v2-spec, Clarification to A4). The strategy is ahead of every one of them in both periods.

† A factor index's figures from before it went live are NSE's back-test of its own rules, not returns anyone could have earned. NSE's launch dates could not be read from an NSE document, so each live part starts at the first date NSE printed an open for the index. Nifty200 Momentum 30 is live from 12 Oct 2020, and over those 5.9 years the strategy made 36.4% a year against the index's 17.5%. Nifty500 Value 50 is live from 16 Dec 2024, and over those 1.8 years the strategy made 7.0% a year against the index's 5.9%, which is too short to judge; the first 1.8 of its 3.6 held-back years are back-test. The other three factor indices are live over the whole period.

What backs it up. The figures in this section come from the run's own output files in data/backtest/v1_final. The measured ones are listed with their definition, file and key in baseline_report.md and its .json; the held-back worst falls above are worked out from the run's nav.csv by the same rule as the whole period's. The method behind each check is explained in research/strategy/v1-verification.md, written on 23 September for the previous run, whose figures it still shows.

• Two engines agree exactly. The production engine and an independent re-implementation written from the specification alone (jobs/crosscheck_v1.py) hold the same stocks with the same ranks on all 31 decision dates, 310 positions in all, and their daily values agree on all 15 series (the strategy, the equal-weight universe and 13 index series) to within 1e-14 of each other (v1_final_check).
• It beats random picks from the same list. Against 500 random portfolios run through the same engine and rules with one random ranking kept for the whole run, it lands at the 96th percentile: 20 did at least as well, and the median random portfolio made 20.7% a year. Against 500 with a fresh random ranking each quarter it lands at the 98.6th: 7 did as well, and the median random portfolio made 19.3%. The first kind turned over 0.5 times a year, the second 3.95, the strategy 1.96.
• It fell less than the market over the whole period. In the months the Nifty 500 rose, the strategy rose 1.30 times as much on average; in the months it fell, 0.66 times as much. Over rolling twelve-month windows it was ahead of the Nifty 500 84% of the time and of the same universe 72%. That did not hold within the held-back period on its own, where its worst fall, -23.8%, was deeper than the Nifty 500's -18.6%.
• The margin over the same universe is probably real, but not proven. In the held-back period it is 13.35 points a year. A 90% interval from a stationary bootstrap of the paired daily returns runs from -2.68 to +29.23 points, and 8.4% of the resamples put the margin at zero or below.

What it does not show:

• Skill beyond known factors. Regressed on IIM Ahmedabad's published Indian four factors over the 82 months from Mar 2019 to Dec 2025, the last month IIMA has published, the strategy's alpha is 5.74% a year at t = 1.21 (1.17 with a small-sample correction), which is not significant. Its value and momentum tilts are, strongly (t = 4.78 and 3.61), and the four factors explain 72% of its monthly returns' variation. Most of the return is the market plus factors that can be bought more cheaply.
• That what is left is more than a quality tilt. Adding a quality factor built from the archive (v2-spec A5, operating-margin version, the same 82 months) lowers the alpha to 5.04% (t = 1.11). Over the 34 months from Mar 2023, where the factor can use return on capital, the alpha is 11.27% (t = 2.61) without the factor, which on its own would count as significant, and 3.31% (t = 0.74) with it: quality accounts for 71% of the alpha in that window, which is only 2.8 years long.
• Robustness to the search. After 195 logged trials, the deflated Sharpe probability of skill is between 0.838 and 0.995 depending on how the spread across trials is estimated, around a usual bar of 0.95.
• A smooth ride. The worst falls were -33.6% from Jan 2020 to Mar 2020, back to its peak by Nov 2020, and -30.7% from Apr 2022 to Jun 2022, back to its peak by Jul 2023. The deepest that started inside the held-back period was -23.8% from Sep 2024 to Feb 2025, back to its peak by Sep 2025. From the end of 2025 to the open of 18 Sep 2026 it is down 3.3% while the same universe is up 8.7%, and it is still below its Feb 2026 peak, having been 16.5% below it at the worst in Mar 2026.

After Indian costs and tax. From Rs 5 lakh at the open of 15 Feb 2019, with itemised Indian trading costs, an assumed 0.1% slippage a side and capital gains tax by lot at the rates in force at the time (jobs/after_tax.py, v2-spec A6), the strategy would have been worth Rs 28.6 lakh if sold at the open of 18 Sep 2026, 25.8% a year, after paying Rs 4.6 lakh of tax along the way; its worst fall on that basis was -31.7%. A Nifty 500 index fund with a 0.16% expense ratio, held throughout and sold at the end, gives Rs 12.7 lakh, 13.1% a year. Before 6 Sep 2019, the start of the index fund whose costs are used, that fund is hypothetical.

Correction, 23 September 2026. Statistics I published alongside the one-shot result on 22 September were wrong. They were my own work, done after the audited build had finished and never checked by anyone else. The worst was a factor regression that subtracted the risk-free rate twice and turned an insignificant alpha into a nearly significant one, on which I then built a claim that the filters carried the strongest evidence. That claim is withdrawn. The headline statistics are now cross-checked against standard libraries (statsmodels, quantstats, arch) in jobs/libcheck_v1.py, and a second independent review checked the corrections themselves.

What the strategy actually is

At each quarterly date, a few weeks after the results deadline so the numbers are fresh:

Filters (a stock must pass all of them): ordinary equity that traded recently, at least Rs 20 lakh of daily turnover, at least 200 sessions traded in the past year, not a bank, NBFC or insurer, four consecutive quarters of results already published, profitable over those four quarters, and a computable market capitalisation.

Across the 31 decision dates that takes an average of 1,853 ordinary-equity symbols that traded recently down to an average of 903: 237 on the first date, when few companies had four quarters of tagged results yet, up to 1,316, and 1,158 on average through the held-back period.

Ranking (five groups, equally weighted, no tuning): profitability and its stability, revenue and profit growth, earnings and sales yield, 12-month momentum skipping the latest month, and low volatility.

Portfolio: the best 10, at most 3 from any sector, equal weighted, a holding kept while it stays inside the top 20, 0.20% cost charged per side, dividends credited as cash.

Full definitions, including every ambiguity resolved during the build and the date it was resolved, are in the specification.

How it was checked

• Two engines. One production engine and one independent re-implementation, written from the specification without sight of each other. They agree on every stock, every date and every return.
• Four audits, for look-ahead, survivorship, trade accounting and statistics. They raised 17 findings and an independent sceptic had to reproduce each one before it was accepted. 16 were real and were fixed, including share counts that were wrong by 100x, a lender filter that used today's list of lenders, and companies that changed ticker being treated as dead.
• A cross-check against published data. Our momentum factor correlates 0.72 to 0.81 with IIM Ahmedabad's published Indian factor returns.
• Selection by rule. 36 portfolio variants were run in-sample, and the plateau rule picked one before the held-back data was unlocked. It picked the configuration that was written down first.
• 198 trials are logged in state/trials.jsonl, including every failure below: 195 when v1 was measured, v2's two runs and v3's one run since.

What came before, and failed

Kept because they are the reason the result above is worth anything.

• The multibagger study found nothing. Low-margin small caps looked like they tripled twice as often. The base rate was wrong, overlapping windows inflated the statistics threefold, and the effect turned out to be dispersion: those stocks halve more often too. Write-up.
• A filing study was void: the orders category turned out to be about a quarter regulatory penalties, carrying the opposite sign.
• Volume spikes may predict negative returns, but an audit re-measurement found almost no effect and the disagreement is unresolved. Write-up.
• Only about 42.6% of randomly chosen liquid stocks beat the Nifty 500 over twelve months, with an honest range of 38.5% to 46.7%. That, not zero, is the bar any strategy has to clear.

---

Layout

Path	What's in it
zen/data/	bhavcopy and filings download, normalisation, DuckDB and Parquet store
zen/monitor/	news collection, ranking, archive analytics, the news to data bridge
zen/signals/	screens and strategies
zen/validation/	look-ahead detection, event studies, the trial log
zen/notify/	email rendering, charts, SMTP delivery
jobs/	entry points that GitHub Actions calls
research/	exploration only, never scheduled

Anything that has to run on a schedule is a module. Notebooks are for looking, not for running.

Workflow	When	What it does
daily_prices.yml	13:30 UTC, weekdays	fetch bhavcopy, store, commit
daily_brief.yml	01:30 UTC, weekdays	build and send the brief
signal_check.yml	after the price update	evaluate strategies, alert if triggered

Signal evaluation is chained off the price update rather than given its own cron, so strategies never run against a stale archive.

---

Data

Source	Coverage	Cost
NSE bhavcopy	daily OHLCV, everything listed including later-delisted	free
NSE corporate filings	every announcement, timestamped by the exchange	free
NSE XBRL financials	quarterly income statement and balance sheet	free
NSE indices	daily levels of the index series counted above, and total-return series for the eight benchmarks in the result	free
NSE insider trading disclosures	every trade promoters, directors and key managers reported since November 2015	free
NSE credit rating filings	the action in each rating letter, read by fixed rules, since October 2025	free
NSDL FPI reports	foreign investors' holdings and net investment by sector, every fortnight since April 2022	free
NSE index lists	the constituents of 16 NSE indices, each version stored from when zen first read it, since October 2026	free
RSS and a news API	Indian market, macro and global news	free tiers
Google Gemini	plain-English explanations, optional	free tier

Getting the fundamentals took longer than it should have. I wrote off the older NSE endpoint as empty, twice, and it wasn't. It returns nothing unless you pass period=Quarterly. With that one parameter it returns 87,449 filings going back to late 2016, and the ones usable for screening start in 2018. That mistake cost about two weeks.

A stock's real news is filed with the exchange, not written by a journalist. The filing shows up within minutes of a board approving it. A newspaper covers it days later, if at all.

Every filing carries an_dt, the moment NSE published it. Anything published after the 15:30 close belongs to the next session, not the one that just ended. session_date is the first session the stock actually traded after that, which is the one the filing could affect. Getting that backwards is a one-day look-ahead, the same shape of error that produced the fake 30% CAGR.

Two thirds of the filings feed is statutory noise. "Copy of Newspaper Publication" alone was 409 of 1,943 filings across four sessions. Routine compliance gets matched and discarded first, so a newspaper advert announcing results can't be mistaken for the results.

One category gets special handling. When the exchange formally asks a company to explain an unusual move, that query arrives after the close, so it confirms the move rather than explaining it. The brief says that, instead of presenting it as a cause.

NSE changed the bhavcopy format on 8 July 2024. Both layouts are parsed into one schema.

---

Running it

python -m venv .venv
.venv/Scripts/activate            # Windows
pip install -r requirements.txt

python -m jobs.update_prices --days 7      # fetch recent sessions
python -m jobs.daily_brief --dry-run       # render locally, send nothing
python -m pytest tests/                    # includes the leak detector
python -m jobs.readme_numbers --check      # README figures still match their sources

Backfill history:

python -m jobs.update_prices --start 2015-01-01 --end 2026-09-07

Secrets never get committed. Locally they go in .env, which is gitignored, and there's an .env.example to copy. On GitHub they live under Settings, Secrets and variables, Actions: SMTP_HOST, SMTP_PORT, SMTP_USER, SMTP_PASSWORD (a Gmail App Password, never the account password), MAIL_TO, and optionally GEMINI_API_KEY and MARKETAUX_KEY.

---

Mistakes

I keep these in public because they're more useful than the findings. Every one produced a confident, wrong answer before I caught it.

What broke	What it did	How I found it
Control group wasn't liquidity-matched	The baseline was 6.4x more liquid than the event stocks, so a size effect looked like a signal. An apparent +20pp edge dropped to +6.2pp and five categories turned out to be noise	Went back over how the control was built
Same-day split and bonus	Only one factor got applied, a 5x price error on DELPHIFX	A single-session return that couldn't be real
Leak detector only truncated prices	A strategy could read tomorrow's filings and still pass	Read the detector against the actual table list
Filing categoriser matched substrings	order missed "orders" and "expansion" missed "expands". The most important filing in the case study that started this got binned as other	Hand-checked a case I already knew the answer to
XBRL parser read a segment context	One division's revenue came through as the whole company's, because segment contexts share the parent's period	Fields came back null where I could see figures in the document
Backfill skipped quarters on restart	171 filings became permanent holes that no rerun could fill	Diffed written rows against the index
Session re-warm keyed on a counter	Never fired once. Throughput fell from 1.8 docs/sec to 0.09 with no error in the log	Watched a live run stall
Connection reuse turned off	My own fix for the last bug added a TLS handshake to every request. 0.42s per document against 0.12s reusing a connection	Timed both against the same URL
Categorising filings by regex on free text	The orders bucket held 14,955 filings. Only about 24% were real order wins and about 27% were regulatory penalties, the opposite sign. The mix also shifted between periods, so my headline finding compared two different things. NSE labels every filing itself, and on that basis the correct bucket is 3,220	Checked NSE's own filing subtypes
Joining filings to prices on equality	trade_date is a calendar weekday, which need not be a session the stock traded. 70,506 filings were dropped, skewed towards those filed before long weekends	Counted the join both ways
Adding a column to the data but not the schema	session_date went into 783,510 rows and not into the CREATE TABLE. Both daily jobs failed for four days. The same shape had shipped a week earlier in another table

The trial log in state/trials.jsonl records the hypotheses I've formally tested, including the ones I abandoned, because how many I tried decides whether the best result means anything. It doesn't yet include about 50 exploratory statistics from piloting the news-reaction study, so the true count is higher than the line count.

---

Where it's up to

Working: the price archive, corporate filings, quarterly fundamentals, both emails, the news to data bridge, corporate-action adjustment, matched controls, the trial log, look-ahead detection, and schema-parity tests that stop the data and the database definitions drifting apart. Strategy v1 is built, tested on its held-back years, re-run on the repaired archive and measured, as above.

Strategy v2 was tested and did not replace v1. v2 added my own hard filters (ROCE at least 10%, debt to equity below 1.5, market cap above Rs 100 crore, P/E at most 70), twelve positions, a graded cash filter, three-tranche entry, volatility sizing and NSE sector caps, all fixed in research/strategy/v2-spec.md before it was run. Two engines, one written from the spec without reading the other, built it and agree on every holding, trade and daily value; the rules and code were public on GitHub before the one full-period run. It passed two of the five bars set in advance and failed three (data/backtest/v2_verdict.md, every figure with its file and key):

	v1	v2
Growth a year	30.7%	21.3%
Worst fall	-33.6%	-26.3%
Sortino, Calmar	1.94, 0.94	1.74, 0.83
IIMA four-factor alpha a year, t	5.7%, 1.21	1.1%, 0.23
After costs and tax, a year	25.8%	17.8%

It made the ride smoother, as intended, but gave up too much return to do it, so v1 stays. Buying every stock that passed v2's filters in equal amounts made 23.8% a year, more than v2's own ranking and sizing on top of them: the filters picked a strong pool, and the extra machinery cost return. The technical entry rule it tested (variant b) bought slightly cheaper but ended lower and was not adopted. v2 is weaker evidence than v1 by construction: its rules were written after v1's results were known, so it had no unseen data to be tested on.

Strategy v3 was tested and did not replace v1 either. v3 is purely fundamental: no price trend and no volatility input anywhere. It buys profitable companies (trailing profit of at least Rs 25 crore, interest cover of at least 2.5, debt to equity below 1.2, P/E of 60 or less, eight unbroken quarters of results), ranks them on accelerating growth, widening margins and growth at a reasonable price, holds eighteen with at most four a sector, never sells because a rank fell, and trims above a P/E of 85. Every rule and every test was fixed in research/strategy/v3-spec.md before any v3 result existed. Two engines, one sharing none of this repository's code, agree on every holding, fill and daily value; they, the leak test and the ten dated clarifications the build needed were public here before the one full run; and a separate verification recomputed all 967 figures of the verdict and checked every exit, trim and fill against the rules (data/backtest/v3/verification.md). It passed one of the five bars (data/backtest/v3_verdict.md). Both columns are 17 February 2020 to 18 September 2026, v1 replayed on the same frozen archive:

	v1	v3
Growth a year	32.1%	24.1%
Worst fall	-32.9%	-44.5%
Sortino, Calmar	1.92, 1.00	1.31, 0.55
IIMA four-factor alpha a year, t	3.7%, 0.80	-0.6%, -0.10
Turnover a year	1.97x	0.52x

The test named in advance, alpha after the four factors and a quality factor, came to -2.5% a year at t = -0.49: no evidence of skill. After costs and tax v3 made 22.1% a year against 13.2% for a Nifty 500 index fund.

What it did show is where the result came from. The filters picked a good pool: every eligible stock held in equal amounts made 26.9% a year, against 25.5% for v1's wider universe. The score ordered that pool: across all eligible stocks its rank correlation with the next period's return averaged 0.05 (t = 3.05), and the top fifth beat the bottom fifth by 22 points a year before costs. But eighteen names bought and held until an exit rule fired did no better than eighteen drawn at random from the same list: v3 sits at the 42nd and 53rd percentile of 500 random fill orders in each mode. A signal that weak needs many more holdings, and a book that re-ranks, to show. v3 was also fully invested on 17 February 2020, weeks before the market's fall. Like v2 it is weaker evidence than v1: its rules were written without looking at a return, but these years had already been seen once, through v1. The spec's section 16 lists its other limits: a window that includes a small and mid cap boom, no balance sheets before September 2022, today's sector labels applied to the past, and whole quarters missing for some companies.

Next: tracking v1 forward, on data that does not exist yet, is the only evidence still to be had.

Limits I'd rather state than have found:

• Income statements go back to 2018, balance sheets only to Sep 2022, because that is when SEBI's half-yearly requirement started producing tagged data. I can screen on leverage and return on capital today but cannot backtest them properly yet.
• NSE only began labelling filing types on 2024-09-23, and the change was a single day rather than a phase-in. Before it, the exchange's own classification does not exist, so a filing study reaching back further is measuring a labelling convention. Each category records its first trustworthy year in code (USABLE_FROM in zen/data/filing_types.py). Studies have to apply it themselves, and not all of them do yet.
• Promoter pledge percentages are not machine-readable here. Only 34 filings carry a figure, so that limit is a manual check rather than a filter.

The momentum screen currently wired in is a calibration baseline, not a strategy for real money. Momentum is well documented and I know roughly what it should return, so it tests whether the harness reports honestly. If validation comes back with an implausible number, the harness is broken. Neither v1 nor v2 is wired into the emails.