Find a signal, then prove it holds (2026)

Minerva is built for exactly this job: the trial ledger feeds a deflated Sharpe whose N is an eigenvalue-based effective trial count, so correlated genetic and Bayesian trials collapse toward one instead of being counted naively, and the family is then passed through PBO/CSCV, Hansen SPA, White Reality Check, Romano-Wolf StepM, Model Confidence Set and BH/BY FDR before a five-gate seal issues a verdict. The search that generates those trials is genuinely large, with grid, genetic and TuRBO Bayesian over a 4,000-candidate pool, a three-stage multi-fidelity scheduler with early stopping and a cloud fleet running roughly 40,000 backtests per optimization. Discovery is strong but not top of the set: a 64-model library plus an AI copilot that compiles natural language into a typed IR and a real backtestable signal, yet nothing that invents novel strategy structures the way StrategyQuant's genetic programming does, so you still arrive with a hypothesis. What it leaves undone is everything after the verdict and around the edges: paper trading against Alpaca IEX plus notification-only signals with explicitly no broker execution, US equities and ETFs only over roughly 2016 to 2026, single-symbol simulation with no shared-capital portfolio engine, market orders only, and no API, SDK or data import, so a surviving signal cannot be scripted, ported or traded from the platform. Weighted mean is 8.45, reported as 8.5.

StrategyQuant X is the only tool in this set built around the first half of this job rather than the second: you arrive with building blocks, not an idea, and the Builder breeds novel strategies for you through random generation plus island-model genetic evolution, with a plain-English AI Assistant added in build 143 that emits a complete strategy you can send straight to AlgoWizard and backtest. Search throughput is real and shipped (vendor claims tens of thousands of strategies reviewed per hour, brute-force and genetic optimizers, sequential optimization, and a cheap-to-expensive cross-check funnel that dismisses candidates before the costly tests run), though there is no model-based Bayesian search and the promised grid computing across multiple machines was postponed and never shipped. The gap sits exactly where this job is decided: the software counts and displays how many strategies it generated but corrects nothing for that count, and no deflated Sharpe, PBO, multiple-testing correction, reality check or purged cross-validation appears anywhere in the features page, the robustness-test docs, the cross-checks docs or the public roadmap. System Parameter Permutation medians over one strategy's own parameter sweep rather than over the mined population, and the correlation filter that would discount redundant trials is a community .sxp plugin you must import and compile yourself, so it does not count. You get the best candidate-generation engine on the market wired to a deep robustness battery, then are handed the survivor with no statistical price tag on the million backtests it beat.

QuantConnect now covers the front half of this job better than its reputation suggests: a documented team of predefined AI assistants runs an Ideas Assistant that scans news, blogs and the dataset catalogue to write up to three testable strategy briefs per run, a Research Validation Assistant that runs ADF, Breusch-Godfrey, Breusch-Pagan and a train/test population stability index in a notebook, and a Backtest Assistant that writes, compiles, fixes and reruns the algorithm until it executes end to end, so you can arrive with nothing and leave with a real backtest. Once something survives, deployment is the best in the field, with the same code going to 20-plus brokerages including Interactive Brokers, Schwab, Binance, Bloomberg EMSX and raw FIX. Where it fails this job specifically is the crux: the optimizer ships grid search and nothing else, parallelism caps at 12 cloud nodes, and the Research Guide merely counts your backtests, parameters and hours and stamps them "Likely Not Overfit" or "Probably Overfitting" without ever converting that count into a deflated Sharpe or a PBO. The Probabilistic Sharpe Ratio printed on every backtest adjusts for non-normal returns against a benchmark, not for the number of variants you burned getting there, and QuantConnect's own overfitting article presents the Deflated Sharpe Ratio purely as reading material.

Wealth-Lab is one of the few platforms here that will invent a candidate for you rather than wait for one: the Strategy Genetic Evolver ships in the base Windows product, randomly generates Building Block strategies from a gene pool, keeps and mutates the top performers across configurable generations, and hands back an editable, directly backtestable, directly tradable strategy rather than a snippet. Around it sit a pre-built strategy library and clonable community Published Strategies, so the front half of this job is genuinely covered. The back half is not. Search is grid, random and shrinking window on local CPU cores, with Bayesian and SMAC available only through the third-party finantic.Optimizer at $89.95, and there is no trial accounting anywhere in the product: the ScoreCard roster stops at plain Sharpe and Sortino, and the Evolver will run thousands of strategy trials and report the winner's raw fitness with no penalty for the count. You get walk-forward with a WFO Efficiency ratio, parameter stability surfaces and a $49.95 Monte Carlo extension to sanity-check a survivor, then a strong path to live broker execution, but nothing that tells you the winner is only the luckiest of ten thousand.

VectorBT is built for the second half of this job and not the first: you must arrive with the idea, because nothing in the vendor documentation generates a candidate strategy for you. What it does have is the best trial-accounting default in this set, a deflated Sharpe whose nb_trials argument silently takes the width of your parameter sweep and whose variance term is computed across all trials, so the multiple-testing correction falls out of the search loop instead of depending on you remembering to count. Search power is the product's identity: grid plus random subset sampling, lazy parameter grids, Numba and Rust kernels, and documented Ray cluster scaling that the vendor pitches as testing billions of combinations, though there is no Bayesian or model-based search and no early stopping anywhere. Once something survives you are on your own, since there is no live trading, no paper trading, and no alerting, and the Alpaca and Interactive Brokers connectors pull data only.

MetaTrader 5 is built for the second half of this job, not the first. It gives you a no-code way to assemble candidates (the MQL5 Wizard turns 20 shipped signal modules plus trailing and money-management modules into a compilable, testable EA, and the new MetaEditor AI Assistant writes MQL5 from a plain-language description), plus one of the cheapest large search capacities anywhere: full grid enumeration, a built-in genetic optimizer, local and LAN agent farms, and the MQL5 Cloud Network at roughly $0.08 per agent-hour, which MetaQuotes says turns months of optimization into hours. Where it collapses is exactly the crux: nothing in the platform knows or cares how many variants you burned. The optimizer reports a table of passes and 2D/3D plots of the result surface, but the Sharpe on the winning pass is raw with the risk-free rate hardcoded to zero, the statistics API stops at STAT_SHARPE_RATIO and STAT_COMPLEX_CRITERION with no deflated Sharpe, PBO or multiple-testing statistic, and the only overfitting guard is a single forward split whose best passes you then cherry-pick, which is selection on the out-of-sample segment. Use it to test hard and deploy instantly to any of 1,000-plus brokers, but do the trial accounting somewhere else.

TradeStation is built for the other job: you arrive with a rule set written in EasyLanguage, and it tests, optimizes and trades it. Its own optimization documentation states the position plainly, that optimization enhances a trading idea and cannot be used to develop one, and the platform lives up to that. For finding a candidate it gives you a Scanner wizard, RadarScreen ranking and a set of insertable indicator-based strategy components, all of which you assemble yourself, with no strategy generator and no AI builder that emits a backtestable strategy (TradeStation MCP explicitly analyzes accounts and previews orders rather than generating or backtesting strategies). Where it genuinely helps is the second half of the job: exhaustive plus genetic optimization feeding a real Walk-Forward Optimizer with Monte Carlo and cluster analysis, then one-checkbox automated execution at the integrated broker. What it leaves undone is the crux, trial accounting: the WFO FAQ tells you to run 1,000 iterations and 5,000 to 10,000 cluster tests and then applies an unadjusted t-Test and a plain Sharpe to whatever won, so searching harder inflates your verdict and nothing in the platform discounts it.

TrendSpider is stronger on the finding half of this job than its no-code reputation suggests: it ships a curated strategy library plus the ML Quant Lab, where a GPT assistant turns a plain-English idea into model inputs, Random Forest or KNN models learn fixed risk/reward signals, and crossbreeding mixes the inputs of good parent models to produce new candidates you did not specify. It is far weaker on the knowing-whether-it-holds half. The Group Strategy Tester multiplies trials up to 53 by default and several hundred on higher tiers across strategies, symbols and timeframes, yet no documented feature counts those trials or corrects the winner for them: no deflated Sharpe, no PBO, no false-discovery adjustment, no p-value anywhere in the results panel, and the sweep cannot even vary indicator parameter values. Out-of-sample discipline exists only as written advice to carve out a date range by hand, so the reward for searching harder is an inflated best result the platform will report at face value, though anything that does survive deploys to a live position-aware bot and, via SignalStack or TradersPost, to a real broker in a few clicks.

Composer is the best tool in this set at the two ends of the job and the worst in the middle. Finding a candidate is genuinely easy: a documented GPT-4 natural-language editor emits a complete, immediately backtestable symphony rather than a code snippet you debug, and the Discover surface plus a roughly 2,780-strategy Symphony Database hands you ready-made candidates with out-of-sample columns attached, and once something survives you deploy it to real money in the same product because Composer is the broker-dealer. What it leaves undone is the entire validation half of the job: its own API reference and its own Create and Backtest Basics articles document no parameter sweep, no optimization, no walk-forward, no Monte Carlo and no cross-validation, so exploring variants means manually re-running by hand. There is no notion of trial count anywhere, and its flagship discovery surface is a sortable leaderboard of thousands of community strategies with no disclosure of how many were screened and no multiple-testing correction, which is an efficient selection-bias machine pointed straight at the user. Use it to find and to deploy, and validate the survivor somewhere else, because Composer's own overfitting page answers this job with six pieces of manual advice instead of a feature.

Trade Ideas is unusually good at the first half of this job and close to useless at the second. It hands you candidates rather than making you arrive with one: a curated bullish/bearish/neutral strategy library you can load and send straight to Backtest Strategy, a July 2026 agentic AI that configures a real running scanner from a plain-English description, and Holly's nightly signals off 60-plus strategies. But the moment you ask whether a candidate actually holds, the platform has nothing: OddsMaker runs one backtest at a time with no sweep or batch mode, the Optimization Tab is a post-hoc filter-segmentation report of a single run rather than a search, and the documented workflow is to tighten filters and re-run until the numbers improve. There is no trial count, no Sharpe ratio of any kind and therefore no deflated or probabilistic variant, no PBO, no out-of-sample split, and roughly 64 trading days of 1-minute history makes even a hand-rolled holdout impractical, so the vendor's own stated defense against overfitting is to go paper trade afterwards. Deployment is the one genuinely strong link, with Brokerage Plus auto-executing surviving strategies live at Interactive Brokers, Alpaca, TradeStation, or E*TRADE, which in practice means it will happily route real money into a setup validated on three months of data.