Backtesting Engine
Test the thesis, and keep weak evidence visible
A backtest replays a strategy against historical market data and reports what would have happened, under stated assumptions about costs and fills. Stretus surfaces risk, drawdown, trade quality and comparison evidence alongside return, and applies a configured gate that refuses promotion when the evidence does not meet the declared objective.
Software infrastructure only. No investment advice, no brokerage services, no guaranteed returns.
01 · Client experience
Broker digital channels
02 · The governed layer
Stretus AI Strategy Infrastructure
03 · Execution authority
Broker execution environment
What is wrong with how backtests are usually presented?
They lead with a headline return, which is the least informative number in the report. A 32% return with a 28% maximum drawdown and a Sharpe of 0.9 is a worse candidate than a 20% return with a 9% drawdown and a Sharpe of 1.7, and a platform that shows only the first figure has actively misled the reader.
For a broker this is not an aesthetic preference. A client who deploys on a headline number, endures a drawdown nobody showed them, and abandons the strategy at the worst possible moment is a support ticket, a complaint, and eventually a churned account.
Illustrative comparison, higher return is not the stronger candidate
| Candidate A | Candidate B | |
|---|---|---|
| Net return | 32% | 20% |
| Maximum drawdown | 28% | 9% |
| Sharpe | 0.9 | 1.7 |
| Review outcome | Do not promote, drawdown burden and risk-adjusted evidence miss the objective | Retain for review, and this is not approval for live execution |
Illustrative data only; not Stretus or customer performance. Candidate decisions depend on the declared objective, configured gates, data quality and independent review.
What does the engine actually measure?
Four families: return, risk-adjusted performance, drawdown and downside, and trade quality. A report that contains only the first family is not interpretable, and the platform is built so that it never appears alone.
One clarification worth making before the detail, because the shorthand "backtest with costs" is easy to misread: this refers to the trading charges a real order would have incurred, not to a fee for running a backtest. Backtesting is part of the platform, and Stretus publishes no pricing of any kind.
Those trading charges are applied at fill time rather than subtracted at the end. For Indian equities that means brokerage, exchange transaction charges, STT, stamp duty, SEBI turnover fees and GST, several of which are charged differently on a buy than on a sell. Modelling them as a flat percentage is convenient and wrong in a way that makes frequent-trading strategies look much better than they are.
It matters because a system that computes results from price differences and applies charges afterwards will have made its sizing and exit decisions on numbers that never existed.
The analytics set
| Family | Metrics |
|---|---|
| Return | Total and net return, annualised return, average outcome per trade |
| Risk-adjusted | Sharpe, Sortino, Calmar, and strategy grading context |
| Drawdown and downside | Maximum drawdown, duration, recovery, downside distribution, adverse outcomes |
| Trade quality | Win rate, profit factor, trade count, holding period, loss streak |
How does the rejection gate work?
A candidate that fails its configured threshold is refused promotion, and the refusal is recorded as strategy-test evidence. A weak result is not hidden, retried silently, or presented as a near miss. It is the outcome of the test.
This is the feature most likely to be argued with internally and the one most worth keeping. A platform that only ever surfaces improvements trains its users to believe every strategy is nearly ready. A platform that says "this candidate failed: win rate below the required threshold" teaches the opposite, and the second lesson is the one that keeps client capital intact.
What are the three improvement modes?
Manual Improve requests a specific change and measures the difference. Loop Improve compares bounded rounds against a declared objective. Target Moves uses selected historical moves to construct a hypothesis, explicitly not proof of generalisation.
Each mode preserves candidate evidence, so a promotion decision can be reconstructed later: what was compared, against what objective, and why the promoted version won. That record is what turns a strategy library into something a compliance function can review rather than a pile of parameters.
Target Moves carries the strongest caveat by design. Constructing a hypothesis from selected historical moves is a legitimate research technique and a direct route to overfitting if the result is treated as evidence. Candidates from it still require independent validation and paper observation.
How results should be read
- High return is not the same as strong risk-adjusted performance
- Win rate alone does not establish profitability
- Historical performance does not predict future outcomes
- A strategy with very few trades has told you almost nothing, whatever its return
Ownership boundaries
Where Stretus sits in the stack
Benefits
Business benefits
Evidence a reviewer can act on
Return, risk-adjusted metrics, drawdown and trade quality in one report, so an approval decision rests on the full picture rather than one number.
Costs modelled, not mentioned
Indian statutory charges are applied at fill time, including the ones charged differently on a buy than on a sell, so the reported figures describe something achievable.
Refusals as first-class output
The rejection gate records why a candidate did not proceed, which is the evidence trail an approval workflow needs and the lesson a client needs.
Comparable versions
Candidate evidence is preserved across rounds, so a promotion can be explained months later rather than defended from memory.
Use cases
Enterprise operating situations
Illustrative operating situations. Availability varies by broker, exchange, account, connector and tenant.
- Challenge
- Review strategy versions, risk inputs and deployment state with a clearer audit trail.
- Stretus role
- Preserve candidate evidence, apply configured gates, and record promotion and rejection context.
- Outcome
- A reviewable history of what was tested, what was refused, and on what grounds.
- Challenge
- Avoid approving strategies whose evidence nobody examined beyond the headline figure.
- Stretus role
- Surface drawdown and risk-adjusted evidence in the same view as return, and gate on the declared objective.
- Outcome
- Approval decisions that can be defended, with the refused candidates on record too.
Security posture
Security considerations
- Test results are tenant-scoped
- Backtest runs, candidate evidence and comparison history sit inside the tenant boundary under role-based access, not in a shared pool.
- No result is an approval
- Passing a gate makes a candidate eligible for review. Live deployment is a separate, policy-gated action that a backtest cannot trigger.
Answers
Frequently asked questions
Which risk-adjusted metrics are reported?
Sharpe, Sortino and Calmar, with strategy grading context, alongside maximum drawdown, drawdown duration and recovery, and the downside distribution.
Does a good backtest mean the strategy will work?
No. It means the rules had an edge in the period tested, under the stated cost and fill assumptions. Persistence is a separate question no backtest can answer, which is why paper observation sits between testing and live execution.
Are transaction costs included?
Yes, and they are applied at fill time rather than subtracted at the end. For Indian equities that includes brokerage, exchange transaction charges, STT, stamp duty, SEBI turnover fees and GST.
What stops a user tuning a strategy until it looks good?
Nothing stops iteration, and nothing should. What the platform does is bound the comparison, preserve the evidence from every round, and require an independent review before promotion. So an overfitted candidate is visible as one rather than arriving as a finished result.
Ecosystem
Related capabilities
AI Strategy Builder
Natural-language strategy creation with deterministic validation: intent becomes a structured, reviewable rule set your client can read before testing.
Risk Governance
Approval workflows, exposure controls, order limits and audit evidence, risk enforced by the platform rather than left to intention.
Monitoring & Observability
Strategy health, order monitoring, latency, exceptions and connected audit evidence, retained for at least five years and identifying the actual user.
Arrange a working demonstration
Review strategy creation, F&O contract handling, backtesting, broker controls and integration boundaries with the team. If you would rather talk to an engineer than a salesperson, say so and we will arrange that instead.
Risk and disclosure
Trading and derivatives involve risk of loss. AI output requires review. Backtests and simulations do not predict future results; live outcomes can differ because of costs, latency, slippage, liquidity, rejections, broker rules and market conditions. Availability varies by broker, exchange, account, connector and tenant. Product information only; not investment advice.