What is Vault?
Vault is a quantitative research engine for Indian markets. It generates a complete, pre-defined universe of systematic strategies, simulates each one on minute-level market data under realistic execution, and validates every result through a series of statistical guards. Each strategy that survives receives a trust score.
On top of that sits an intelligence layer. It analyses all the results jointly, finds strategies that combine well, tracks live performance against historical expectations, and answers your questions through an AI agent.
You don't write code or tune parameters. You describe what you want, and Vault returns validated strategies with the evidence behind them.
Why one strategy tells you very little
Optimise enough parameters on one idea and some configuration will always produce a strong equity curve. Most of that performance is variance, not edge. This is data snooping, and no amount of analysis on a single result can separate signal from noise.
Vault takes the opposite approach. It evaluates the full parameter space under one consistent methodology, then analyses the results jointly. Cross-sectional analysis reveals three things a single result cannot:
- Robustness: does performance hold across neighbouring parameters, or collapse outside one point?
- Regime dependence: is the edge stable across market regimes, or confined to one?
- Correlation: are the returns uncorrelated with other strategies, so they can be combined into a more efficient portfolio?
Vault is built in two layers:
- The vault: a pre-defined strategy space, where every strategy is validated through statistical guards and assigned a trust score.
- The intelligence layer: cross-sectional analysis of those results, and an agent that turns it into answers.
Part 1: Inside Vault
Strategies are generated, not hand-written
Strategies in Vault are not written one by one. They are defined as strategy families. Each family exposes a parameter set: entry time, exit time, stop-loss, profit target, strike selection, re-entry rules, and market-condition filters.
Vault expands every family into the full Cartesian product of its valid parameter combinations, with invalid or contradictory configurations pruned.
The complete universe is fixed before a single result is observed. This is a small detail with large consequences. Adding strategies after observing results biases the search toward historical artefacts, and every downstream statistic, from p-values to rankings, becomes silently invalid. A pre-registered universe keeps the hypothesis count known and the search honest.
Execution under realistic constraints
Every strategy is simulated on minute-level market data with a conservative execution model. Statutory and exchange charges are applied at the rate schedule in force on each trade date, fills are only permitted in minutes with recorded traded volume, and price series are cross-validated against a secondary data source before any simulation runs.
Fair sizing. If position size is a function of option premium, low-priced contracts receive disproportionately large positions. A ₹8 option can end up with 20 times the lots of a ₹150 option, and the ranking ends up rewarding leverage instead of skill. Vault evaluates every strategy under two sizing regimes, a fixed 1 lot and notional-value sizing, so no strategy ranks highly on hidden leverage.
Every run is persisted as a complete record: parameter vector, trade log, daily return series, and per-regime performance attribution.
The guards
A strong result is where the work starts, not where it ends. Every strategy passes through six guards, applied in sequence so that each stage only evaluates candidates that survived the previous one.
- Cost guard: is the edge positive under adverse cost and slippage assumptions?
- Sizing guard: is performance consistent under fixed-lot and notional sizing?
- Plateau guard: is the result supported by its parameter neighbourhood, or is it an isolated spike?
- Noise guard: does it beat the maximum statistic produced by the same search on permuted data?
- Regime guard: does it hold up under current market structure, not only in past regimes?
- Holdout guard: does it pass on a locked out-of-sample period, evaluated exactly once?
Each guard targets a specific failure mode of historical simulation.
Cost guard
Every strategy is evaluated under three cost scenarios: optimistic, central, and adverse. The adverse scenario deliberately widens slippage and inflates charges, because the true bid-ask spread on every historical fill is unobservable, and assuming otherwise would bias results upward.
If the edge is only positive in the optimistic scenario, the strategy is rejected. Its expected return never exceeded its transaction costs.
This guard is highly selective. A 1-lot position collecting ₹8 of premium pays roughly 8% of that premium in charges before generating any return. Many strategies that appear profitable are simply under-costed.
Sizing guard
A strategy must rank well under both sizing regimes. If it only performs at large notional size, the return is driven by leverage. If it only performs at 1 lot, it is unlikely to scale to real capital. The divergence between the two results is recorded and surfaced to you.
Plateau guard
If a 9:30 entry is profitable but 9:25 and 9:35 are not, the result is not a strategy. It is an overfit point estimate.
Vault scores each strategy against its parameter neighbourhood, the strategies whose parameter vectors differ by a single step. A strategy ranks well only when its neighbourhood is consistently profitable. A sharp peak surrounded by losses ranks below a broad, stable plateau, even when the peak has a higher standalone metric.
There is a subtle trap here that most tools ignore. Distinct parameter vectors frequently collapse onto identical trade sequences. A range of premium targets, for example, can all select the same contract every day. Counted separately, those configurations would inflate the apparent evidence. In reality they are a single observation.
Every strategy therefore receives a trade fingerprint: a hash of its realised trade sequence, covering entry timestamp, instrument, quantity, and exit reason. Strategies with identical fingerprints are deduplicated. Plateau width is measured in distinct trade histories, not in parameter rows.
Noise guard
This guard addresses the central problem of large-scale search: multiple hypothesis testing.
When a very large number of strategies are evaluated, some will look exceptional purely by chance. The relevant question is not whether the best strategy looks good. It is whether it beats the best strategy the same search would find if the data contained no exploitable structure.
To estimate that, Vault runs a stratified permutation test. Each trading day is kept intact while the sequence of days is shuffled, with permutations stratified by weekday and days-to-expiry. The entire search, including winner selection, is then re-run on each permuted dataset, 500 times over.
This produces an empirical null distribution of the maximum statistic. A strategy must clearly exceed the upper tail of that distribution, not merely beat zero. If the search consistently finds strong "strategies" in permuted data, the pipeline itself is flagged as leaking information, and the campaign is halted.
The scope of this test is stated plainly. It shows that a strategy outperforms what the same search produces when the inter-day structure of the market is destroyed. It does not prove the edge will persist. No test can.
Regime guard
Market microstructure changes. Lot sizes are revised. Expiry days move. Some weekly contracts are discontinued entirely. A strategy that was valid under a prior regime may not be executable under the current one.
Vault tags every trading day with the structural regime in force and reports performance per regime. It never averages a strong 2021 with a weak 2025 into a single blended figure. It also distinguishes two failure states:
- Decayed: the strategy remains executable, but its edge has degraded.
- Untradeable: the contract or rule it depended on no longer exists.
Holdout guard
Each time out-of-sample data is inspected and the strategy is adjusted, that data leaks into the fitting process. After a few iterations, the out-of-sample test is effectively in-sample.
Vault locks the most recent market period as a holdout set. Each strategy is evaluated on it exactly once, and that evaluation is recorded in an append-only ledger. Any subsequent evaluation request is refused.
The current regime is also short. With only a few hundred trading days under today's rules, statistical power is limited and only large effect sizes can be confirmed. When the sample cannot support a conclusion, Vault reports "insufficient evidence" rather than a confident-looking point estimate.
The trust score
Every strategy that survives the guards is assigned a trust score from 0 to 100.
The score aggregates the guard outputs: margin over adverse costs, agreement between sizing regimes, plateau width, distance above the null distribution, and current-regime performance. It is also corrected for the total number of hypotheses tested. The larger the search, the stronger the evidence required to earn the same score.
The trust score is fully decomposable. You can open it and inspect every guard result behind it. A strategy with a score of 71 might clear the cost and noise guards comfortably while having limited evidence in the current regime, and that breakdown is shown explicitly.
Most strategies don't make it
Guard thresholds are never relaxed to increase the survivor count. If a guard rejects 90% of a family, that is information about the market, not a sign the guard is too strict.
A campaign with zero survivors is still a valid outcome. It prevents the same lesson from being learned with real capital.
Part 2: Intelligence on top
One strategy describes one strategy. An entire strategy space, evaluated under a single methodology, describes the market. This layer is where better strategies come from: not from testing one more idea, but from analysing all the results jointly.
Behaviour clusters
Strategies that differ in their parameters often behave identically in practice. Vault clusters strategies on realised behaviour: trade-day overlap, return correlation, and response to common market events.
A large cluster is usually one underlying idea with minor parameter variation. That matters in two ways:
- You see ideas, not duplicates. Instead of many near-identical strategies, you see one representative and the size of the cluster supporting it.
- Robustness is measured at the idea level. An idea that holds across an entire cluster is stronger evidence than a single configuration that happens to stand out.
Combining strategies
The strongest risk-adjusted results rarely come from a single strategy. They come from strategies whose weaknesses offset each other.
Because every strategy in Vault has its daily return series stored on a common calendar, Vault can compute the full correlation structure between any set of them. A strategy that earns steadily in low-volatility, range-bound sessions and one that earns on large directional moves may each look unremarkable alone. Combined, their drawdowns rarely coincide, and the portfolio has a smoother equity curve and a higher Sharpe ratio than either component.
Vault searches for these combinations across clusters, not within them. Combining two variants of the same idea adds no diversification, because their returns are highly correlated. Combining distinct ideas that each survived the guards, and that perform in different regimes, is where the real improvement lies.
Every combination is held to the same standard as a single strategy. Each component must have survived the guards independently, and the combined portfolio is validated on the holdout period, so a combination cannot be fitted to the past any more than a single strategy can.
Regime awareness
Because every strategy is scored per regime, Vault knows far more than a point-estimate return. It knows that a strategy performs well in low-volatility, range-bound conditions and poorly around large overnight gaps. It knows which strategies depended on a weekly contract that has been discontinued. It knows how many trading days of evidence exist under the current regime.
This changes the answer you get. "This strategy returned 14% a year" becomes "this strategy performed well from 2021 to 2023, degraded after the 2024 rule changes, and has about 150 trading days of evidence under the current regime." It also informs which strategies to combine: ones whose strong regimes cover each other's weak ones.
Live drift tracking
Every historical result is an implicit forecast, and Vault tests that forecast.
Strategies in Vault continue running on live market data as it arrives. Their live performance is compared against the distribution their history predicted. This is the cleanest out-of-sample test available, because the data did not exist at evaluation time, so nothing could have been fitted to it.
When a strategy's live performance drifts outside its expected distribution, its trust score is downgraded and Vault notifies you. Trust is not assigned once. It is re-earned every trading day.
The agent
On top of all this sits an AI agent you interact with in plain language.
You describe what you want: "A low-drawdown options strategy for NIFTY that doesn't hold overnight." The agent queries Vault across guard results, clusters, regime breakdowns and live tracking, and returns strategies and combinations that held up. It then explains them: what each one does, the conditions it tends to perform in, where it tends to fail, and why it earned its trust score.
The agent does not generate numbers. Every figure it quotes is retrieved from a stored record that you can open and verify.
You can also refine a strategy conversationally, for example a tighter stop or a later entry. The agent does not search for the single optimal parameter, because that would reintroduce overfitting. It only moves within the strategy's plateau, the region of neighbouring parameters that all passed the guards.