Backtest · NVDA
Backtest NVDA — Candle by Candle
NVDA is the hardest stock in this catalogue to backtest honestly, and the reason is not volatility. It is that the archive contains two different companies under one ticker. Before roughly 2023 NVDA traded like a cyclical semiconductor: it followed the memory and GPU cycle, and it drew down 50%+ twice. After 2023 it traded like an index component that moves the index. A rule fitted to the second era and validated on the whole file will look far better than it is.
The regime split is the whole problem, and it is testable
Split your sample deliberately rather than letting the backtest average two regimes into one number. Run your rules on 2018–2022 and on 2023–now as separate tests and compare the two win rates. If they disagree badly, you have not found an edge on NVDA — you have found an edge on one market condition that NVDA happened to be in.
This is not an NVDA quirk, it is just the clearest case of it in the catalogue, which is why it is worth learning here. The same discipline is what out-of-sample testing is for.
The second thing that will distort your numbers is earnings. NVDA's report has become a scheduled event that moves the whole index, and the overnight gap that follows it is routinely larger than a normal week's range. Four sessions a year can therefore carry more P&L than the other 248 combined. Decide before you start whether your strategy holds through earnings or flattens into them, write it down, and apply it to every occurrence — including the ones where holding would have won.
What the NVDA data actually is
The archive runs from May 2018 to the most recent session we hold, on every timeframe from one minute to monthly. It is open from the first chart — there is no plan that unlocks more of it, and no shorter window for people who have not paid.
Equity data here is regular hours only — 09:30 to 16:00 New York, no pre-market and no after-hours. That is not a gap in the data, it is what an equity session is, and it has a consequence worth planning for: every trading day opens with a gap rather than continuing from the previous close, so any rule you carried over from an index future or an FX pair is about to meet a discontinuity it has never been tested against.
Prices are split-adjusted throughout, so a chart of 2019 shows what you would compare against today rather than the raw tape. Getting this wrong is the classic silent equity-backtest bug: an unadjusted series prints a 90% single-day crash on the split date, and a stop-loss rule tested through it produces numbers that mean nothing.
How to backtest NVDA without fooling yourself
Open a 15-minute chart, step forward, and mark every setup your written rules would have taken. Do not skip the four earnings gaps; log them as they come, with whatever your rule says to do about them.
Log the outcome in R rather than dollars. NVDA's price has ranged from single digits to hundreds across this archive on a split-adjusted basis, so dollar results are not comparable between 2019 and now and R results are. Size each test with the position size calculator so the risk is constant even when the price is not.
Then run the same rules on AMD and MU. Seven of the seventeen names here are semiconductors, and they move together — so a rule that "worked on three chips" may be one move counted three times. Testing a non-semi like NFLX is the honest control.
Frequently asked questions
Is NVDA backtesting free?
Yes, and there is no card and no trial timer. You get NVIDIA back to May 2018, with every tool the product has, for 2,000 candles of replay — about two hours of stepping through a chart. After that you top up for $5, or pay $19.99 once and never think about it again. Nothing you buy expires. US equities used to be the one part of the catalogue behind a plan. They are not any more.
Is the NVDA data split-adjusted?
Yes, throughout — including the 10-for-1 split in June 2024. An unadjusted series would print a 90% single-day fall on the split date, which any stop-loss rule tested through it would read as a catastrophic loss that never happened.
Does the data include pre-market and after-hours?
No. It is regular-session only, 09:30 to 16:00 New York. That is deliberate: it is the session the vast majority of volume trades in, and it means every day opens with a genuine gap rather than a synthetic continuation.
How far back does NVDA go?
May 2018, on every timeframe from one minute to monthly. That window covers the 2018 semiconductor drawdown, the 2020 crash and recovery, the 2022 bear market and the AI expansion — four genuinely different regimes, which is the minimum a stock backtest needs.
Can I backtest NVDA earnings gaps specifically?
Yes, and it is one of the more useful things to do here. Step to the session before a report, take the position your rule dictates, then advance one candle into the open and record the result. Do that for every earnings date in the archive and you will have an actual distribution instead of a memory of the good ones.
Open NVDA and start stepping
Real NVIDIA history, split-adjusted, one candle at a time. No account, no card, 2,000 candles free.
Open the chart →