Helvetic Research turns the AI interface you already use into a quantitative research desk. Describe a strategy, portfolio, signal, or track record; your AI translates the request into a validated specification and calls a deterministic research engine. The result comes back with metrics, robustness diagnostics, provenance, integrity checks, and reproducibility metadata.
The hosted service exposes 90 MCP research tools. It does not place trades, connect to brokers, or provide investment advice.
Why it exists
Asking an AI to write and run a one-off backtest can produce plausible output without a stable method, consistent annualisation, realistic costs, or an audit trail. Helvetic separates the two jobs:
flowchart LR
A[Research question] --> B[Your AI interface]
B --> C[Validated JSON strategy specification]
C --> D[Deterministic Helvetic engine]
D --> E[Research result]
E --> F[Metrics and charts]
E --> G[Robustness tests]
E --> H[Provenance and audit bundle]
The AI handles conversation and tool selection. Helvetic handles calculation, validation, storage, provenance, and reporting. Arbitrary client code is never executed inside the hosted web/MCP process.
Research → Validate → Understand
What you can research
Strategy inputs
- Plain English or pseudo-code
- Pine-like / TradingView strategy logic
- Deterministic structured JSON
- Strategy IR v3 typed JSON AST
- Client-supplied OHLCV bars, including intraday data
- Client-computed signal or position series
- Multi-asset prices and target weights
- An external return stream or fund track record
Markets and portfolio work
- Equities, ETFs, indices through suitable tickers or ETF proxies, FX, commodities, crypto, futures proxies, and macro/reference series
- Single-instrument rules, long/short research, rotation, multi-asset portfolios, synthetic options, universes, and custom supplied books
- CHF, USD, EUR, and GBP reporting currencies
- Monthly, quarterly, semi-annual, annual, and buy-and-hold portfolio rebalancing
- Efficient-frontier, maximum-Sharpe, minimum-volatility, and risk-parity analysis
Signals and costs
- 40+ indicator families with parameterised periods: moving averages, RSI, MACD, Bollinger Bands, ATR, ADX, Ichimoku, Supertrend, volume profile, pivots, Fibonacci, Heikin-Ashi, Aroon, Vortex, KST, and more
- Nested AND/OR conditions, crosses, rolling levels, calendar rules, higher-timeframe filters, stateful exits, trailing stops, position age, and sizing rules
- Commissions, spreads, slippage, volatility-linked slippage, market impact, short borrow, futures roll, crypto funding, FX conversion, management fees, custody fees, and TER
See the full, public capability catalogue.
Robustness and statistical diagnostics
Helvetic is designed to challenge a result, not merely produce one:
- Walk-forward out-of-sample analysis
- Monte Carlo simulation with IID and block bootstrap
- Parameter sweeps and cost/signal-lag sensitivity
- Probability of Backtest Overfitting via CSCV
- Probabilistic and Deflated Sharpe Ratio
- Historical stress periods and custom scenarios
- Volatility, trend, and macro regime analysis
- Factor exposure, alpha decay, capacity, and implementation-cost analysis
- EVT tail-risk analysis, VaR, and Expected Shortfall
- Trade-order permutation testing to separate a real edge from sequencing luck
- Benchmark-regime Sharpe decomposition showing which regimes the edge actually comes from
- A 12-component Research Grade measuring strength of evidence, not whether a strategy is suitable to trade
Due diligence and record-keeping
Beyond running your own ideas, Helvetic helps you vet others' claims and track what you actually did:
- Track-record reconciliation re-simulates a stated set of rules and reconciles the claimed return, CAGR, Sharpe, and drawdown against a clean re-run, field by field, flagging any drift beyond tolerance. The due-diligence answer to "is this track record real?".
- Duplicate-strategy detection fingerprints a strategy by the trades it actually makes and flags near-duplicates among your saved runs, so a "new" idea that is really an old bet in a different costume is caught even when the rules look different.
- Research journal lets you log a trade you actually took with your rationale and a strategy tag, close it for an automatic P&L, return, and R-multiple, and review a per-strategy scorecard (win rate, expectancy, profit factor). It closes the research-to-reality loop. This is record-keeping only: Helvetic executes nothing and gives no advice.
Outputs and reproducibility
Every completed run records the information needed to understand what was actually tested:
- Data provider and retrieval timestamp
- Adjustment method and coverage information
- Missing-data, stale-cache, and provider-disagreement warnings
- SHA-256 dataset and strategy fingerprints
- Engine version, assumptions, configured costs, and integrity checks
- Frequency-correct performance and risk metrics
Results can be viewed on the interactive web result page or inside supported AI chats through an MCP App. They can also be exported as branded PDF, Excel, CSV, JSON, or a self-verifying audit bundle containing raw series, formulas, hashes, assumptions, and provenance.
Connect an AI client
Helvetic is a hosted, remote MCP server. There is nothing to install: you point your AI client at one URL and authorise it once with Google.
-
Visit helveticresearch.com and continue with Google.
-
Add the hosted MCP endpoint to your client and complete the OAuth prompt:
https://helveticresearch.com/mcp- Claude (Desktop or web): Settings → Connectors → Add custom connector → paste the URL.
- ChatGPT (Developer mode / connectors): add an MCP server with the URL above.
- Cursor / Codex / other MCP clients: add a remote MCP server pointing at the URL.
- Full, client-specific steps with screenshots: /connect.
-
Send a complete research request. A good first prompt names a ticker, a date range, the rule, and the costs:
Backtest SPY from 2005-01-01 to 2026-01-01. Go long when price crosses above its 200-day simple moving average and exit when it crosses below. Apply 10 bps transaction costs, show the key metrics, then run a walk-forward test and an overfitting check.
More copy-paste starter prompts and ready-to-run strategy specifications are in
examples/.
The Free plan includes 250 research credits per month and full available history from community data sources. Heavy workflows use more credits than lightweight lookups.
Data coverage
Built-in intraday retrieval uses Yahoo/yfinance and preserves full OHLCV timestamps. Practical
provider lookbacks are approximately 7 days for 1m, 60 days for 2m–90m, and 730 days for
1h plus resampled 2h/3h/4h. Tiingo and Coinbase fallbacks are daily-only. Longer or
licensed intraday histories can be supplied through custom OHLCV tools.
Community data is useful for research and prototyping but is not a licensed point-in-time institutional feed. Historical constituents, delistings, symbol changes, corporate actions, futures rolls, and survivorship controls can be incomplete. Read the full data coverage and limitations.
Important boundaries
- Research only: no orders, brokerage connections, or portfolio execution
- No investment advice: outputs are analytical and educational
- Synthetic options: Black-Scholes simulations, not historical option-chain fills
- No arbitrary code execution: strategies run through supported deterministic schemas
- Private by default: research records are account-scoped; publication is explicit
- Failures remain visible: degraded data and unsupported requests return warnings or clear errors rather than silent success
Public resources
- Explore a real sample result
- Read the methodology
- Review security controls
- Read the research disclaimer
- Browse the website
- Report vulnerabilities privately using SECURITY.md
About this repository
The production research engine is a hosted service and is not open source. This public repository provides the canonical project overview, capability documentation, data limitations, security policy, and registry metadata. An open-source verifier for exported Helvetic audit bundles is planned for this repository.
Questions and product feedback: [email protected]