What does Fidelity Lab actually do?

You upload your own fills. It re-nets every leg against the venue's published fee schedule, runs 11 contamination detectors over the file, picks the ruler that matches each market family, freezes a rule on a time-ordered TRAIN split and evaluates it on held-out TEST, and applies Benjamini-Hochberg across every rule you submitted. It returns SURVIVES, DISSOLVES or UNTESTABLE.

How is this different from a backtester?

A backtester runs your strategy and gives you a number. Fidelity Lab grades a result you already have and is allowed to refuse. It is a second opinion, not a replacement for QuantConnect, TradingView or your own code.

Do you see or keep my strategy?

You send legs, not code. Your rows are stored in an isolated customer-plane database on a separate server, under a database user with no grant to anything else, and are never joined to or used to train anything on the operator's own trading data.

Free (1 report/month), Starter $49/mo or $490/yr, Pro $199/mo or $1,990/yr, Enterprise $499/mo or $4,990/yr. Annual billing is two months free.

TradeHouse · strategy validation

Fidelity Lab

You send us your trades. We tell you whether the edge is real — the same way we tell ourselves. Point-in-time walk-forward, cost-netted from the venue's own fee schedule, with a contamination report that names the specific artefact and the specific legs.

Run a free report See pricing

The only validator that is allowed to say no. Every other tool returns a number. Ours returns UNTESTABLE when your book cannot answer the question — and tells you exactly which column would make it answerable.

Three answers, and one of them is a refusal

A verdict you can act on has to be able to come back negative. All three of these are real outputs.

SURVIVESA rule frozen on the training window still pays after costs on data it never saw, and clears Benjamini-Hochberg across everything you submitted.

DISSOLVESThe in-sample edge does not reproduce out of sample, or it does not survive the venue's real fees. The report says which of the two killed it.

UNTESTABLEThe book cannot answer the question — too few independent days, no entry-time signal to freeze, or an unrecognised market family. We refuse to grade it rather than invent a number.

What we check, and where each check came from

Eleven contamination detectors. Every one of them exists because it caught us first, on our own book — each has a planted failure case that must trip and a healthy control that must stay clear.

Detector	Severity	What it catches
Lookahead	BLOCKER	An outcome or a feature stamped before entry, or a record created after the outcome it predicts.
Oracle lag	MATERIAL	Settled win-rate far above the price paid — a stale quote scored the fill. On our own kalshi book this one turned an apparent +$3,406 gross into −$7,355 net once the flagged cells were removed.
Crossed book	MATERIAL	ask < bid, dead rows, and outcome sets that are not mutually exclusive and collectively exhaustive.
Signed-PnL direction	BLOCKER	NO-side legs scored with the YES convention — a sign bug that manufactures profit.
Hindsight selection	MATERIAL	The cohort was chosen because it was already profitable; in-sample rank fails to predict out-of-sample rank.
Mark tautology	BLOCKER	The exit price is the contemporaneous mid, so the P&L is an accounting identity, not a trade.
Unrealised as P&L	MATERIAL	Open positions and markout rows summed as if they were realised.
Zero fee	MATERIAL	A zero fee charged on a venue the schedule prices as non-zero.
Duplicate legs	MATERIAL	Identical (timestamp, ticker, side, size, price) rows inflating the sample.
Degenerate pin	ADVISORY	Binary contracts entered at ~0 or ~1, where the apparent hit-rate is structural.
Span adequacy	MATERIAL	The whole submission lives in too few independent days to support any conclusion.

The right ruler for the market, or none at all

One metric for every market is how a binary book gets graded with an instrument built for continuous prices. Fidelity Lab routes by market family — and returns UNRULED and refuses rather than grading a family it has no ruler for.

predictions

Closing-line value

Direction-corrected CLV against a de-vigged sharp closing line — computed only from a close_fair column you supply. We do not redistribute a line; if you don't have one, the report says CLV is unmeasurable, which is not the same as zero.

equities & spot

Arrival slippage + IC

Implementation shortfall against your arrival VWAP/mid, plus the information coefficient of your signal. Needs an arrival_vwap or arrival_mid column from you; without it the slippage field is reported null, not guessed.

options

Variance premium

Entry implied vol against subsequently realised vol. Needs entry_iv and realized_vol from your own records.

perps & futures

Basis + funding carry

Carry decomposition from basis and funding_rate, so a funding-harvest book is not graded as if it were directional.

every family

Real fee schedule

We ignore the fee in your file and re-net from a per-venue schedule, labelling each leg SOT_MODELED, SOT_REALIZED or FLOOR_UNMODELED. The last one says out loud that we have no model for that venue and applied a 2% floor.

every submission

Adequacy + multiplicity

Day-clustered effective sample size with lag-1..5 autocorrelation inflation, and Benjamini-Hochberg at q=0.05 across every train-frozen rule you evaluated. Send twenty cohorts and the bar moves.

What you'll get

One JSON document, and the same thing rendered as HTML. This is an abridged real report shape — the numbers below are from our own equities book run through the customer path.

// GET /fidelity/report/{run_id}
{
  "verdict": "UNTESTABLE",
  "reason": "n_eff = 3.0 day-clusters against 256 required for this effect size,
             and no train-frozen rule survived on held-out test.",

  "pnl": {
    "as_submitted_usd":        1284.11,
    "after_exclusions_usd":     -311.40,
    "cost_source":  "venue fee schedule (SOT_MODELED 2790 legs, FLOOR_UNMODELED 0)"
  },

  "contamination": [
    { "code":"UNREALIZED_AS_PNL", "severity":"MATERIAL", "legs":5627,
      "detail":"open markout rows summed as realised (-$527.29)" },
    { "code":"ZERO_FEE",          "severity":"MATERIAL", "legs":10651,
      "detail":"fee 0 on a venue the schedule prices non-zero" },
    { "code":"DUPLICATE_LEGS",    "severity":"MATERIAL", "legs":8252,
      "detail":"77.48% share an identical (ts,ticker,side,size,price) tuple" }
  ],

  "ruler": { "family":"equities", "name":"VWAP_SLIPPAGE_AND_IC",
              "information_coefficient": 0.270, "pairs": 2790,
              "slippage_bps": null, "slippage_note":"no arrival_vwap column supplied" },

  "walk_forward": { "split_date":"2026-08-27", "gap_days":1,
                    "threshold_frozen_on":"TRAIN",
                    "test_net": -0.02521, "t": -18.16 },

  "adequacy":  { "n_legs":10651, "n_eff_day_clusters":3.0, "required":256,
                 "binding_constraint":"independent days, not legs" },

  "multiplicity": { "method":"benjamini_hochberg", "q":0.05, "K":2, "survivors":0 },

  "provenance": { "engine":"f4236.1.0", "report_sha256":"…" }
}

Note the shape of that answer: a real information coefficient of +0.270 in sample, and the walk-forward still dissolves. That contrast is the entire product.

Pricing

Annual billing is two months free (16.7% off). Prices in USD, exclusive of tax. No contract, cancel any time.
Fidelity Lab is billed separately from the TradeHouse platform tiers — it grades your own trades and needs no TradeHouse data subscription.

Free

no card required

1 report per month
up to 5,000 legs
all 11 contamination detectors
cost-netted P&L from the fee schedule
full walk-forward verdict

Start free

Starter

$49/mo

or $490/yr — 2 months free

10 reports per month
up to 100,000 legs
CSV and JSON upload
per-market rulers
BH-FDR across every cohort you send
HTML + JSON report, downloadable

Choose Starter or pay annually — $490

MOST CHOSEN

Pro

$199/mo

or $1,990/yr — 2 months free

unlimited reports
up to 2,000,000 legs per run
report API + webhook
scheduled re-runs of a saved book
read-only broker connect — rolling out
priority support

Choose Pro or pay annually — $1,990

Enterprise

$499/mo

or $4,990/yr — 2 months free

everything in Pro, no leg limit
custom per-market ruler built for your book
bring your own closing line or reference feed
analyst review of every verdict
private deployment option
named contact, SLA on turnaround

Talk to us

What is live today, and what is not

Live: CSV and JSON upload of your own legs, all 11 detectors, per-market rulers, fee-schedule cost netting, walk-forward with a TRAIN-frozen threshold, BH-FDR, HTML and JSON reports, and the report API.
Rolling out on Pro: read-only broker and venue connect (Alpaca, Tradier, Kalshi, IBKR, Coinbase). The adapters are built and the scopes are read-only — we will never request an order scope — but connect is not open to self-serve signup yet. Until it is, Pro is billed for unlimited runs and the API, and connect arrives at no extra charge.
We do not sell you a closing line. The CLV ruler runs on a close_fair column you supply. If you don't have one, the report tells you CLV is unmeasurable rather than substituting something we invented.
We do not sell you a price tape. Slippage and IC are computed against your arrival prices. Running a rule over our own market data is a separate, licensed conversation, family by family.
We grade, we do not advise. Fidelity Lab is backward-looking analytics on data you supply. It is not investment advice, it produces no signals, and it executes nothing.

Questions

Is this a backtester?: No, and it is not trying to replace one. QuantConnect, TradingView, AmiBroker and the rest run your strategy and hand you a number. Fidelity Lab takes the result you already have and adversarially checks whether it means anything. Most customers use both.
Why would I pay for a second opinion?: Because the failure mode is expensive and invisible. Every detector on this page fired on our own book first. One of them turned a $3,406 apparent profit into a $7,355 loss.
What do I have to send?: At minimum: entry timestamp, venue, side, size. Everything else unlocks a specific check, and GET /fidelity/schema tells you exactly which column would buy you which detector or ruler before you pay anything.
Where does my data go?: Into an isolated database on a separate server, under a database user with no grant to anything else, keyed to your tenant on every query. It is never joined to, mixed with, or used to train anything on our own trading data.
What if my book is too small to grade?: You get UNTESTABLE and the number of independent days you would need. That is a real answer and it costs you a report credit — we would rather tell you that than sell you a false positive.

TradeHouse provides data and tools, not advice. Fidelity Lab performs backward-looking statistical analysis on data the customer supplies. Nothing here is a personalized recommendation, an offer, or a solicitation to buy or sell any security or contract. Past performance is not indicative of future results (CFTC 4.41(b)). Pattern-A compliance.