Minerva review (2026)

Minerva is the narrowest serious platform on this list and the most careful. Its engine computes the visible-data boundary centrally and hands it to every signal, so a strategy structurally cannot look at data it should not see. Fills are delayed by a documented latency contract that even handles the case where an intraday bar's close and the next bar's open share a timestamp. Stops that gap through fill at the open rather than at a price nobody could have got. Order capacity is capped against prior-session volume for every slippage model, and a symbol with no history gets no fill at all rather than one against invented liquidity. On the four axes that decide whether a backtest number is correct, it scores 8.5, ahead of everything else assessed here.

The price of that is breadth, and the list is long. US equities and ETFs only. Roughly ten years of history, no tick data. Market orders only, which means no limit or bracket orders and no partial fills. Single-symbol simulation, so there is no portfolio engine sharing capital across positions. No live execution: paper trading and notifications. No API, no SDK, no data import, and a closed codebase, which means nothing it claims can be verified from outside.

Disclosure: BacktestScore is operated by the same team as Minerva. Its scores here come from the same published formula as every other platform, and its rigour axes come from a line-by-line audit that is described in the evidence. It ranks second, behind QuantConnect, because breadth and value carry 55% of the formula and Minerva loses on both.

Minerva is a hosted research platform for US equity and ETF strategies. You pick a signal from a library of published models or describe one in natural language, which is compiled into a typed intermediate representation and then into a machine-code signal function. You optimise it on a cloud fleet, and the platform returns a verdict rather than a statistics table: a composite score built from five statistical gates, with a pass or fail seal.

The design goal is narrow and explicit. It is not trying to be the platform that reaches the most markets or executes the most order types. It is trying to make the number it gives you correct, and to tell you when that number is more likely to be luck than edge. Everything about the product follows from that choice, including the limitations.

There is no code editor, no API and no SDK. Strategies live inside the platform. That is a real constraint if you want to port a survivor elsewhere, and it is the main reason the value score here is low relative to free alternatives.

The engine is event-driven and bar-by-bar, compiled ahead of time so the hot path carries no interpreter overhead, with a pure-Python reference implementation held in lockstep and tested against it per model.

Three things stand out against the rest of this list. The boundary of visible data is computed by the engine and injected into every signal kernel, rather than each strategy slicing its own window; that removes an entire class of accidental look-ahead. The execution latency policy is a first-class contract carrying data, decision and broker latency, and it explicitly refuses a zero-latency fill on contiguous intraday bars where one bar's close timestamp equals the next bar's open. And purge in cross-validation is derived from the model's declared lifecycle horizon plus an observed maximum holding period, rather than being a fixed guess.

Costs are modelled rather than approximated: a square-root market impact model with separate temporary and permanent components, participation capacity computed from prior completed sessions and enforced for every slippage model, and broker commissions with itemised regulatory fees. A symbol with no prior volume history gets zero capacity, so the trade is skipped instead of filled against liquidity that never existed.

What the engine does not model: limit and stop-limit orders, partial fills, queue position, and any intrabar path, since the finest resolution is one minute. For strategies whose behaviour depends on where inside a bar the price went, that is a real limit.

Bars are bundled, US-listed equities and ETFs, roughly April 2016 to June 2026, at resolutions from one minute to daily. There is no tick data and no way to bring your own.

The handling is careful in ways that matter more than the size of the archive. Prices are split-adjusted at storage and dividend-adjusted using a CRSP-style backward cumulative method from a verified reference table, and crucially the adjustment state and the reference version are part of the bar cache key. That means a long-running process cannot serve bars adjusted under one manifest alongside bars adjusted under another, which is a subtle corruption most systems never guard against.

Universe membership is point-in-time, and choosing to apply today's members to the past stamps a survivorship-bias warning onto the result rather than silently proceeding. Delisted tickers are retained.

The honest weakness is coverage. Ten years of one national market is thin for any test of regime stability, and the absence of tick data caps what the execution model can claim. Compare QuantConnect, which ships US equity data back to 1998 at tick resolution across ten asset classes.

Researcher is $99 a month, or $79 a month billed annually. Trader is $249 a month, or $199 annually, and adds the portfolio engine, paper trading and streaming decisions across up to 200 portfolios. Institutional is quote-based. A 14-day money-back guarantee applies.

The part that changes the arithmetic is credits. Compute is metered at $0.70 per credit on top of the subscription, and a serious optimisation campaign runs tens of thousands of backtests. Budget for the subscription plus usage, not the subscription alone.

Against the field this is expensive. Backtrader, Backtesting.py, Zipline Reloaded and NautilusTrader are free. VectorBT's paid tier is a fraction of this. AmiBroker is a one-time licence. What you are paying for is the validation layer and the bundled data, and whether that is worth it depends entirely on whether you would otherwise build the statistics yourself, which most people do not.

It fits someone who has been burned by a backtest that looked good and traded badly, who works in US equities, and who wants the statistical machinery without building it. If you have ever tested two hundred variants and kept the best one without adjusting for having looked two hundred times, this is the platform built around that specific failure.

It does not fit you if you trade futures, forex, crypto or options; if you need limit orders or intrabar precision; if you want to execute live from the same platform; if you need more than ten years of history; or if you want to script anything, since there is no API and no export path for a strategy.

A reasonable pairing is to research here and execute elsewhere, which is what the absence of live trading forces anyway.