Chapter 13: The Due-Diligence Playbook

This is the chapter I wish I could hand to every person who has ever been pitched an AI trading product. Think of it as the professional version of "trust, but verify," minus the trust part. Verify first. Then decide whether trust is warranted.

Everything here applies equally to a $50-a-month bot, a signal service, a copy-trading portfolio, or your own homebuilt system. The questions don't care about the price tag.

The ten questions

Ask any AI trading vendor these, in writing, and watch what happens. Legitimate operations answer crisply. Marketing operations change the subject.

Is the track record third-party audited? Show me monthly returns, including the losing months, not screenshots of winning trades. Screenshots are advertising. Audits are evidence.
What are the Sharpe ratio, maximum drawdown, and longest recovery time? (Definitions below.) Anyone selling performance who can't quote these doesn't understand their own product.
Is the backtest survivorship-free and point-in-time? Were delisted and bankrupt companies included in the historical universe? Was only information available at the time used? If they can't explain these terms, their backtest is fiction.
How many strategy variants were tested before this one was presented? Fifty variants tested means the "winning" backtest had a 92% chance of looking good by luck alone. Multiple testing without correction is the most common lie in the industry.
Are results net of ALL costs? Commissions, spreads, slippage, data fees, subscriptions, borrow costs. Gross returns are a fantasy novel. Net returns are the bank statement.
Who are the named humans, and what are their regulatory registrations? Look them up on Investor.gov. "Our team prefers privacy" is not an answer.
Can I withdraw at any time? Any lock-up, advance fee, or withdrawal gate is disqualifying. Your money must remain your money.
What happens in a stress event? Kill switch? Position limits? Overnight exposure caps? Flash-crash behavior? "Our AI handles it" is not an answer.
Does the vendor trade its own capital alongside clients, and how is the vendor compensated? You want to know exactly whose interests align with yours, and whose don't.
What data does the AI actually use, and can you describe the methodology without hand-waving? "Proprietary AI" with zero disclosure is marketing. Every legitimate quant can explain their approach at a conceptual level without revealing trade secrets.

Reading a track record like a professional

Three numbers matter more than the headline return:

Sharpe ratio, excess return per unit of volatility. Rough guide: 1–2 is a good professional quant result. Good systematic equity strategies often run 0.5–1.0. Above 3, assume overfitting or error until proven otherwise. Below zero, you're paying for the privilege of losing.
Maximum drawdown, the worst peak-to-trough loss. A strategy that made 40% but once lost 35% getting there is a different animal than the headline suggests. The MAR ratio (annualized return divided by max drawdown) compresses this into one number: higher is better, and anything under 1.0 means the worst loss exceeded the average annual gain.
Recovery time, how long the worst drawdown took to heal. A 30% drawdown that recovered in six months is adversity. One that took four years is a career event.

And the bias that corrupts all of it: survivorship. You're only shown what survived, the winning variant, the fund that didn't close, the trader who didn't blow up. "Losers don't get screenshotted" should be engraved above every trading forum.

One more statistical subtlety worth knowing: significance comes from calendar span, not data frequency. Two years of five-minute bars sounds like a lot of data, but it's still two years of market regimes. It takes a much higher Sharpe to clear statistical significance than twenty years of daily data does. A vendor bragging about "millions of data points" over eighteen months is selling you resolution, not reliability.

The fee autopsy

Add up everything from Chapter 5's iceberg, then do the division the vendors hope you'll skip. A $5-a-month flat fee on a $300 account is a 20%-per-year drag before the strategy earns a penny. Percentage fees scale fairly. Flat fees punish small accounts brutally. Subscriptions, spreads, slippage, data, compute, taxes (next chapter): the strategy doesn't need to beat the market. It needs to beat the market by more than all of this combined. Most don't. That isn't pessimism. It's arithmetic.

Suitability: the question vendors never ask

A fiduciary adviser must ask whether an investment suits you: your goals, timeline, risk tolerance, tax situation. No bot, signal service, or copy-trading platform does this, because none of them owes you that duty. Leverage, options, concentration, and drawdown tolerance interact with your personal situation in ways no algorithm knows.

So ask it yourself, or pay someone who will: does this strategy's worst case fit inside my life? If a 30% drawdown in the satellite account means delaying retirement or missing tuition, the strategy doesn't suit you no matter how good the backtest looks. Suitability isn't a compliance checkbox. It's the difference between a setback and a catastrophe.

Next: the silent partner in every trade, the tax code, and why high-turnover AI strategies carry a second, invisible hurdle rate.