Sixty ideas, run until they broke

A chart, on a screen

It started as a Flask dashboard scanning NSE equities for CPR, Camarilla, relative strength, reversals and fractal coils. The organising idea was a timeframe ladder: weekly reads yearly levels, daily reads monthly, four-hour reads weekly, and everything intraday reads the day.

The first data came from Yahoo Finance, and nine of those first 512 files were not Indian equities at all: crude, gold, copper, natural gas, palladium, platinum, silver, corn, wheat. There was a multi-asset scope, and it got dropped. Narrowing to NSE equity and index options was a deliberate early call, and the abandoned commodity leg is the evidence of it.

Building the ground truth

Upstox OAuth, the Nifty 500 matched against the instrument master, and a hive-partitioned parquet store: daily candles back to 2000, minute candles from 2022.

The detail that mattered was in the cleaner. Session detection became coverage-based and era-relative: a date counts as a real trading day only if at least a fifth of that year’s active symbols traded. Before that fix it had been silently discarding pre-2010 days, because it was judging 2004 by how many symbols trade in 2025. Rows are removed; values are never touched.

Four strategies, three engine generations

Camarilla long and short, CPR long and short, confluence long and short, across six timeframes, through three generations of engine.

The acceptance checklist was written down before anything ran: expectancy-per-risk at or above 0.10, at least 300 trades, profit factor 1.25, at least 40% of symbols profitable, drawdown no worse than 20R. Settled design decisions were logged as “do not revert” with the reason attached to each.

The deployable

Walk-forward logistic scoring ranked trades by win probability, and it worked: out-of-sample, the top quartile hit 75–92% win rate against 42–44% for the bottom. Futures were modelled with real lot sizes and margin. Options were priced off actual bhavcopy premium rather than a formula. A live scanner ran on the Upstox REST API.

The headline was a Sharpe of 3.88, a maximum drawdown of −5.7%, and a book that survived 2008 down only 4%.

Every number in that paragraph is wrong. The next entry is why.

The teardown

Entry was booked at the level. But the signal only confirms once a bar closes beyond that level: on daily data, about 1.95% beyond it. You cannot fill at a price the bar has already left. The entire modelled edge was living inside that gap.

Flip between the two and watch every sleeve cross the line:

The signal is a lagging report that the move already happened.

Then we tried to rescue it. A stop order resting at the level fills on the touch, in real time, but it drags in exactly the fakeouts the closing confirmation had been screening out, roughly half of all touches, each losing between 0.6% and 2.7%. Three leading filters were pre-committed before testing: volume on approach, ADX primed, higher-timeframe trend. None of them clears cost. ADX made the daily book actively worse.

Four more biases fell out of the same audit. Survivorship inverted the trust ordering: the universe is the 2024 F&O list projected backwards, so the long history is the most contaminated, and that famous −4% in 2008 is the book largely sitting the year out: thirteen trades against a normal fifty-four to sixty-nine. Capacity was fiction, arithmetic over tens of thousands of signals. Slippage was uncalibrated and charged nothing on entry. And the structure itself (which family, how many sleeves, which exits) had been chosen in-sample.

One thing did survive the teardown, and not as a strategy. If the breakout side is dead because half of all touches fail, then the failures themselves are a population worth watching. That became a tab, and the tab argues with whoever opens it:

In the live Hub, breaks are classified broke → failing → failed / reclaimed. Its own warning sits above the table: the fade tilt on choppy days is real and monotonic, but the mechanical average is about one basis point before costs. Trending → trade breakouts found no support at all.

Honest costs, honest gates

Brokerage is close to zero through the API, so the real costs are securities transaction tax and slippage. We rebuilt the round trip per vehicle from its components: 0.238% for cash delivery, 0.035% for cash intraday, 0.026% for futures. Spread was estimated with Corwin–Schultz, market impact with a square-root law.

Here is the dial itself, computed from the store and playable. Press play and watch five and a half years of participation breathe: the share of all 502 names above their own 50-day average (bright) and 200-day average (faint). Green when the majority is participating, red when it is not.

And the rotation the hub’s sector tab watches, running on the real data. Each dot is a sector’s relative strength against the universe, plotted against the momentum of that strength; the tail is its last eight weeks. Press play and watch 274 weeks of money moving between sectors:

The options detour

Two gigabytes of one-minute option chains across twelve underlyings, including expired contracts pulled through the expired-instruments API. We built the scalp book and ran it forward on paper.

Buying short-dated premium intraday is negative carry. It loses on average, always. What had looked like a momentum fade was entirely expiry-day theta, and cutting by days-to-expiry proved it.

The kill factory

Instead of hand-coding hypotheses one at a time, we built engines whose purpose is to reject them.

A describer that cannot p-hack: ten low-redundancy gauges per minute, z-scored using only the past, with outcomes measured from an entry you could actually have got, feeding a judge that demands a chronological split, trains at or before 2025, tests on held-out 2026, nets off costs, and uses robust statistics rather than means.

Alongside it: a hundred pre-registered hypotheses, none testable without approval. A factor zoo of 462 factors that collapsed to roughly two independent bets: size, and momentum, which turned out to be relative strength wearing a different name at a correlation of +0.98. A reinforcement learner trained on real level geometry against an identical agent trained on placebo twins, where the profitable placebo was the alarm.

Machine learning, astrology, Gann geometry and a TradingView indicator all went through the same bar and all died. The astrology engine was verified against Solar Fire to within 1.4 arcseconds first.

Forensics on four real traders

We took the machine and pointed it at actual trading records. Two thousand two hundred and fifty-eight trades decomposed: skill lands at p = 0.45, and the profit is one name. Another book’s stated rule is dead, while its discretionary selection is real at p ≈ 0.000: the trader was good at something other than the thing they believed they were doing. A +68.3% run over ten days turned out to have an entry that fails four different ways. A VCP playbook sign-flips on every lever when you move it from the F&O list to a wide universe, which is survivorship.

The steward crisis

The integrity checker had gone fifteen days without an update and was covering three of seven stores, because it enumerated what to check. Adding a store broke nothing; it simply went unchecked, forever. Meanwhile it was globbing one filename pattern while 499 symbols carried 55,457 duplicate rows under a different one, and it reported clean. Three orphaned dashboard processes were squatting on ports, one of them twenty-five hours old.

A checker that names what it checks can only find bugs in what it names.

The fix was architectural rather than diligent. A registry file became the declared desired state: every store, every service, every port. The session manager discovers what is actually on disk and running, and fails if anything is not declared. Add a directory and preflight goes red until you declare it.

The Hub’s System tab says when the last refresh failed instead of degrading quietly; every scheduled run carries a chip per step so a failure points at its own cause. The doctor reports 0 fail, 1 warn, 58 pass. The single covered line is the broker account holder’s name.

The second brain

The research corpus is hundreds of documents, and the reason it stays usable is a rule, not a tool: an Obsidian vault sits over the whole repository, and a standing instruction requires every new diagnostic, spec or finding to join the graph connected, never as an orphan node. One hundred and fifty-eight notes, one hundred and ninety links between them.

The live book, and the placebo

A positional momentum book on relative strength went to forward paper trading. Then we tested it against a thousand dart-throwing placebos.

The selection is real: it beats random name-picking. But its Sharpe lands at roughly equal-weight buy-and-hold, which is a much less interesting claim. And the 48% compound headline splits into 85% across 2021–23 against 13.7% across 2024–26.

In the Hub, the paper book prints its assumptions under the chart where they can be argued with: next-open entry, full costs, ten basis points of slippage, idle cash at six percent. The regime gate is off, so the watchlist is what the rules would buy and the position count is zero out of ten. The rupee figures are simulated capital.

The sweep, and what the literature actually offers

We swept the public internet across six lanes looking for mechanical systems and found the same thing everywhere: there is no undiscovered entry signal out there. What the literature genuinely offers is the risk layer.

Ding’s generalisation of Grinold gives the frame: information ratio is capped at IC divided by the volatility of IC, regardless of how much breadth you add. More bets cannot rescue an unstable signal.

Kabaddi: turning the machine on the person

The last thread pointed the instrument at its own operator. Four rounds of interviews: one spoken, twenty-two minutes, transcribed rather than translated so the Hinglish trading vocabulary survived intact; three typed, verbatim. The goal was to write down the discretionary system that actually gets traded, as a specification.

The specification surfaced a reframe that invalidates every mechanical test that came before it. The target is 0.7–0.8% in the underlying, which becomes roughly 10% on the premium, and then the position is closed. The daily level everyone was testing against sits 5–6% away, and is deliberately not the target.

The system never needed the entry to predict direction. It needed the position to visit a small favourable excursion.

That is a question about the path a trade takes, not about where it ends. Every test in the repository up to this point had measured endpoints.

So we pre-registered 104 cells, split train at or before 2023, test across 2024–25, holdout 2026. Cells clearing the null at p95, after cost, after a red-team pass: zero. Best net expectancy: −12.95 basis points per trade. Random entry on the same stock-days beat the best signal by about ten points of compound annual return. Two candidates survived the first pass, were independently re-derived from raw data by separately written code, reproduced exactly — and were then destroyed by attacks that reversed their sign. A clean null, not an inconclusive one.

The durable output is not a signal. It is this, measured on a real account’s actual behaviour: