Appendix

api/risk.py adds portfolio-risk/backend to sys.path and calls RiskService(demo=True). That is the entire Portfolio Risk page. Demo mode because the live path calls input() for a broker token, which blocks a web service on stdin with no way to say so.

Flat api/main.py with no APIRouter, Depends or CORS; raw psycopg2 connection per call; lib/api.ts copied verbatim; the Route Handler proxy pattern that keeps the browser off FastAPI. No code and no data imported — it holds real order data.

That copy deliberately excludes outputs/, so the backtesting pages read AutoTrader/research/outputs directly. Two copies of a result means two answers to what a strategy returned.

Automerge and CI tooling — a watch script and a launchd plist.

research/outputs for backtest runs. Nothing in AutoTrader is modified by this app, and it never touches the 5435 staging cluster the collector writes to.

Live prices during the session

The dashboard reads the SSD archive on port 5434, which holds end-of-day bars. Today's ticks are on the staging cluster (5435), which is the collector's write path during market hours. Keeping the browser off that cluster is why a page load cannot disturb capture.

What you get instead: Everything here is as of the last completed session.

Positions, orders and realised P&L

order-management holds real Zerodha order data and portfolio-risk/backend/accounting can compute P&L from it. Neither is wired in. This app is read-only market data, and the deployment we discussed puts it on a public host.

What you get instead: The risk page runs on a seeded synthetic book, labelled as such. Its maths is the real engine; only the positions are invented.

Point-in-time index membership

index_membership is a reconstruction seeded from an alphabetically-sorted universe. Its current "Nifty 50" omits RELIANCE, HDFCBANK, ICICIBANK, ITC and SBIN while including AARTIIND and AKZOINDIA, which it claims have been members since 1996.

What you get instead: Index filters are current-membership only. "Was X in the Nifty 50 in 2019" cannot be answered from this database.

Bank Nifty, Fin Nifty, Midcap constituents

Membership lists exist for Nifty 50, 100, 200 and 500 and nothing else. Those indices can be charted from index_prices but their members cannot be ranked.

What you get instead: The Banks panel is derived by company name from the Nifty 500 -- 26 names, a superset of Bank Nifty's twelve.

Market capitalisation, free float, listing date

No shares-outstanding column exists in any of the 190 tables, so market cap cannot even be derived. securities.first_seen is populated on 4.7% of rows and means "first seen in index membership", not listed-on.

What you get instead: Turnover is offered instead, labelled traded value rather than size.

Implied volatility

Tested rather than assumed. Put-call parity -- model-free, and the gentlest possible check -- yields implied rates of -2059%, +392%, -181% on NIFTY's own chain, because each contract's daily close is struck at a different moment against a moving underlying. If parity fails that badly, a Black-Scholes inversion would produce plausible nonsense. There is also no risk-free curve anywhere in the schema.

What you get instead: No IV, no greeks from daily data. It would be defensible on derivative_bars_minute (69.9M rows, 2022 onward) and that is future work, not a gap here.

Every return and price level is back-adjusted for splits, bonuses and consolidations.

equity_daily is raw and prev_close is the previous close, not NSE's adjusted one. Across the Nifty 500 over six years there are 127 log-returns beyond ±40% across 110 symbols. Unadjusted, RELIANCE's 5-year volatility reads 38.05% instead of 22.40% and its excess kurtosis 526 instead of 3.3; INFY's maximum drawdown becomes −87.9% dated 2020 instead of −50.7% dated 2026.

Volatility is annualised over 247 trading days, not 252.

This market traded a mean of 246.91 days a year over 2015–2025, range 244–250. The textbook 252 overstates every volatility by about 1.03%.

close, never last_price.

Both are fully populated and differ on 80.3% of rows by a mean 37.5 bps. NSE's own prev_close chains to the prior close 99.4% of the time and to the prior last_price 4.3%. close is the official closing price; last_price is the last trade.

F&O open interest is quoted in shares, and both shares and contracts are shown.

NIFTY 23500 CE held 8,068,840 on 2026-09-09 at lot size 65 — 124,136 contracts. Labelling the raw figure "contracts" overstates it 65× and looks entirely normal.

MCX is converted from paise; CDS is already in rupees.

Two tables with an identical schema and different units. GOLD reads 15,243,000 and means ₹152,430 per 10g; USDINR reads 95.03 and means 95.03. Verified across six MCX contracts and four currency pairs.

Commodity and currency charts use the front month, and name the contract.

GOLD had three contracts trading on 2026-09-09 at 152,430 / 153,983 / 156,256. Grouping by name interleaves them into a sawtooth that is not a price.

An argmin over every listed strike, not the strike with the most open interest.

Σ callOI·max(0,S−K) + Σ putOI·max(0,K−S), minimised. On 2026-09-09 the biggest NIFTY call wall was 23,700 and max pain was 23,600 — the two differ, which is what makes the lazy implementation wrong rather than merely different.

Every percentile ships with the window it was ranked against.

The FII long/short ratio has a structural break: its median ran 1.87 in 2012 and 0.14 in 2026. Today's 0.124 is the 3rd percentile of all history and the 15th of the trailing three years. Only the second is useful.