ACADEMY / Strategy generation

Computer Specs for Program Trading and Backtesting

An order machine only needs to stay up: 4 cores, 16 GB, a small SSD, a UPS and a backup connection are enough. A backtest machine needs many cores, 32 GB or more of RAM and a large SSD, and the larger the number of candidates times the number of bars, the more the wait depends on the processor or graphics card. Buying the two machines separately is cheaper and safer than one machine that does both.

By Soren Huang · Published 2026-11-05 · updated 2026-11-05 · Editorial policy

Contents

What computer does program trading need? Start by separating the order machine from the backtest machine

Think of program trading hardware as two machines. The order machine needs only an ordinary office-class PC with 4 cores, 16 GB of RAM and a small SSD; what matters is that it runs 24 hours without interruption, has a wired connection plus a backup, and sits behind a UPS. The backtest machine is the one that needs a multi-core processor, 32 GB or more of RAM and a large SSD, and a graphics card only for large searches. Buying the two separately costs less and is safer than one machine that does both.

The order machine is the computer that runs your strategy during the session and sends orders to your broker. The backtest machine replays your strategy rules on historical bars and compares many candidate strategies. Compare them to household appliances: the order machine is the refrigerator, low power but never allowed to stop; the backtest machine is the oven, run at full heat while in use and switched off when the job is done.

There are three concrete reasons to buy them separately. First, when a backtest maxes out the processor, the order program on the same computer slows down, and a few extra seconds can appear between the close price and the order going out, which is exactly the moment you cannot afford to be slow. Second, a backtest machine gets new software, restarts and part swaps all the time, while an order machine is best left untouched for a year. Third, an order machine is cheap and uses little power, so you do not mind leaving it on 24 hours a day; a backtest machine is expensive, but you can switch it off when idle and upgrade it when the budget allows. If you are just starting with TAIEX futures program trading, begin with a single order machine; for the research workflow, see Program trading for TAIEX futures: zero to first strategy.

Item	Order machine	Backtest machine
Processor	4 cores is enough	As many cores as your budget allows, or add a graphics card
Memory	16 GB	32 GB or more
Storage	Small SSD	Large SSD
Uptime	On 24 hours a day, no interruptions	Can be switched off when done
Network	Wired first, plus a backup connection	Ordinary home internet
Power	On a UPS	No UPS needed

An order machine and a backtest machine have almost opposite requirements: one must never stop, the other must compute fast.

How many cores and how much RAM does a program trading computer need?

An order machine with 4 cores and 16 GB of RAM is enough. For a backtest machine, cores and RAM should grow with the number of candidates times the number of bars: the longer the data and the more variants you try, the more the core count matters. A core is a unit inside the processor that can work on one task on its own. RAM is the workspace where the computer holds data and intermediate results, like the size of a desk: if the desk is too small, you keep clearing it and fetching things again.

The order machine's work is light: receive quotes, calculate the indicators once when each bar closes, and send an order if the conditions are met. With a charting program and an order program both open, 4 cores and 16 GB are more than enough. What matters more is choosing a model that is low-power, quiet and has few parts, because fewer parts means fewer points of failure.

A backtest machine is different. You can estimate the workload of one study with a simple formula:

Backtest workload ≈ number of candidates × number of bars

Here the number of candidates is how many rule and parameter combinations the study will try, and the number of bars is how long the historical data is that each candidate has to run through. Multiply the two and you get the total the processor has to compute. Core count matters because candidates are independent of each other and can be spread across cores: 16 cores run 16 candidates at the same time, 4 cores run 4. RAM has to hold the whole price series plus the trade records each candidate produces, so 32 GB is enough for a single-market daily study, and for a multi-market minute-bar study you should plan on 64 GB.

Hypothetical example

Suppose one core of a computer takes 0.1 seconds to backtest one candidate on ten years of daily bars (about 2,500 bars). One strategy: 0.1 seconds. 1,024 candidates: about 100 seconds, a cup of coffee. 1 million candidates: about 28 hours on a single core, about 7 hours on 4 cores, and about 1.7 hours on 16 cores. Run the same 1 million candidates on ten years of 1-minute bars, where the bar count is several hundred times larger, and 4 cores need months. This is why someone running a single strategy never notices the core count, while someone running a large search notices it a great deal.

Backtest workload grows as candidates multiply by bars. Moving right (more candidates) and down (finer bars) both make the square larger, and only when both grow does the core count really decide whether you wait hours or days.

Does a GPU help with backtesting?

A GPU helps with large studies that search hundreds of millions of candidates at once. A daily-bar backtest of a single strategy or a small number of candidates runs fine on an ordinary processor, and the NQ daily backtest in this article was run on a CPU. A graphics card suits work where the same small calculation is done on thousands of candidates at the same time; backtesting one candidate is exactly that kind of small calculation, so the card's advantage appears only when the candidates are numerous. With only a few dozen candidates the wait barely differs, and the money is better spent on RAM and an SSD first.

Here is an actual example. AlphaPrime (Valynox) ran a search on the CPU: NQ daily bars, 2019-01-02 to 2025-06-30, long and short, 1 contract, with a US$20 commission and US$10 slippage charged on every trade. The search tried 1,024 candidate strategies and kept one, ranked by return relative to drawdown. The one it kept uses a single ATR (Average True Range) rule for entry and exit, with no stop loss, no profit target and no filter. It is a placeholder rule built from a standard indicator, used to show what a CPU backtest produces, and it is not a test of this article's main subject.

Look at the three cells for number of trades, max drawdown and profit factor, and at the period field: this is the summary of a single-market, daily-bar search over seven and a half years of data, run on the CPU. NQ · daily · 2019-01-02 to 2025-06-30 · CPU engine backtest

Source: AlphaPrime product engine (CPU), run 2026-10-31. NQ daily, 2019-01-02 to 2025-06-30, long and short, 1 contract, commission $20 and slippage $10 per trade.

Number	Result
Net profit (closed trades)	$233,415
Max drawdown	$96,430
Trades	172
Win rate	59.3%
Profit factor	1.52
Average trade	$1,357.06
Time in market	81.97%

Read this table as follows. The 172 trades, $233,415 net profit and $96,430 max drawdown belong to the best-looking of 1,024 candidates, picked after the fact, and picking after the fact by itself makes the numbers look better. Bailey et al. (2014) show that the more variants you try, the more the best backtest is overstated. This does not mean the ATR rule earned that money, and it is not a result you should expect. Its use is to show that a single-market daily search with about a thousand candidates finishes in reasonable time on an ordinary computer, and that a graphics card is not required.

When is a graphics card worth adding? Look at AlphaPrime's published benchmarks: on one RTX 3090, 1,268,200,000 strategy evaluations finished in 18 hours 23 minutes (measured 2026-10-05, TY 37-minute bars, 2015-01-01 to 2026-09-25, 114,371 bars per candidate, with commission and slippage included in every evaluation). At this scale of candidates and bars, how long you wait depends on the processor or graphics card you use, and the benchmark speed also depends on the graphics card model, the data length and the study settings. Whether it runs on the processor or the graphics card, AlphaPrime's GPU backtesting applies the same commission, slippage, session and next-bar fill settings to every candidate, so the hardware changes only the waiting time, not the results you compare.

How big a drive do you need for historical data?

Daily data takes almost no space, minute bars start to run to hundreds of MB, and only tick data makes you plan SSD capacity carefully. Bar size, not the number of markets, decides the drive size. Tick data records every single trade, and one 1-minute bar can contain thousands of trades, so tick data for the same period is several orders of magnitude larger than minute bars.

To see the gap, take three markets that Taiwanese traders commonly trade. NQ trades almost 23 hours a day, which gives more than 1000 minute bars a day; NQ and MNQ futures: Nasdaq-100 specs, sessions and backtest costs covers its sessions and costs. Gold futures also trade nearly around the clock, and Gold futures backtesting: specs, margin and trading hours lists the specs. UMC stock futures trade far shorter hours than NQ, so the same minute bars add up to far fewer per year; specs and contract month changes are in UMC stock futures: specs, code, rolls and backtesting. The daily bars of all three markets combined are still smaller than one year of tick data for a single market.

Hypothetical example

Assume each bar stored as a text file takes about 60 bytes. Ten years of daily bars is about 2,500 bars, under 1 MB. NQ has about 1,380 1-minute bars a day, so ten years is about 3,400,000 bars, roughly 200 MB. For tick data, assume NQ has 1 million trades a day at 30 bytes each: about 30 MB a day, about 7 GB a year and about 70 GB for ten years, and several times that if you store it in a format that includes bid and ask quotes. So for daily and minute bars alone, a 500 GB SSD is already roomy; only multi-market, multi-year tick data needs 2 TB or more, plus a separate backup drive.

Disk needs jump with bar size rather than rising in line with the number of markets. The daily bars of three markets together are smaller than one year of tick data for a single market.

Two habits are more useful than buying a bigger drive. First, keep raw data and research working files apart: raw data is read-only and backed up once, while working files can be deleted and rebuilt. Second, choose an SSD over a hard disk, because a backtest reads the data in one continuous pass, and an SSD cuts the read time from minutes to seconds, which you feel more than extra capacity.

Will you miss orders if the computer shuts down or the network drops?

Yes. When the strategy runs on your own machine, stop loss orders and exit instructions are only calculated and sent by that computer after the close price appears, so the moment it shuts down, crashes or disconnects, those instructions do not exist. That is why the order machine needs a chain of backups: a UPS, a backup connection, and manual exit through your broker's app. A UPS (uninterruptible power supply) is a battery placed in front of the PC and router that takes over the moment mains power stops, keeping the machine up long enough for you to deal with your position or shut down normally.

Think of the backups in order, each layer covering one kind of failure, with the last layer always a person:

Power outage: Plug the PC and the router into the UPS, and the fiber modem too. Protecting only the PC is useless because the network goes down with the power. Pick a capacity that lasts 20 to 30 minutes; the goal is to last until you have exited, not until power returns.
Fiber line down: Set a phone hotspot or a second connection to take over automatically. Unplug the network cable outside trading hours once and confirm that the order program reconnects by itself instead of stopping at an error window.
PC crash or VPS disconnection: Open your broker's app and check open positions and order status. An order that the strategy thinks is still working only counts if you can see it at the broker.
Last layer: Exit with a market order in your broker's app, and call your broker's representative to confirm. Write the representative's phone number and the exit steps on paper and stick it next to the screen, because in a power outage your phone is the only tool that works.

Every kind of failure has its own layer of backup, and the last layer is always a person exiting through the broker. A stop loss order calculated by the strategy exists only while the computer that sends it is still alive.

The chance that your shutdown time overlaps a live session is higher than it feels. TAIEX futures (TX) trade the night session from 15:00 to 05:00 the next day and the day session from 08:45 to 13:45; NQ trades on CME Globex from Sunday to Friday, 17:00 to 16:00 the next day Central Time (CT), with only a one-hour break. In other words, whatever time you shut down, almost always some market you trade is open. Exchanges adjust their hours, so rely on the exchange's contract page. To see each market's current hours converted to Taiwan time, use the free futures market hours checker.

There is another way to move risk off your own computer: if your broker offers server-side touch orders or stop orders, place the protective order on the broker's server right after entry. The stop's existence then no longer depends on your computer staying on, and your own computer handles only entries and normal exits.

Laptop, home PC or trading VPS?

A laptop suits watching the market and light research, a desktop suits heavy backtesting, and a trading VPS suits running orders 24 hours; a common setup among Taiwanese futures traders is one desktop for research plus one VPS for orders. A trading VPS is a virtual computer rented in a data center. You connect to it by remote desktop, and it stays on thanks to the data center's backup power and multiple network paths, so your home computer or network going down does not affect it.

Item	Laptop	Home desktop	Trading VPS
Best for	Watching the market and light research	Heavy backtesting	Placing orders 24 hours a day
Backtesting power	Limited	Strong	Weak
Running orders 24 hours	Not recommended	Possible, but needs a UPS and a backup connection	Yes
Outage and disconnection risk	High	Medium	Low (the data center has backup)
Cost structure	One-time	One-time	Monthly fee
Expandability	Low	High	Add-ons available

Each of the three machines does one thing best: the laptop is convenient, the desktop computes fast, the VPS stays connected. None of them does all three well.

A laptop cannot run orders for 24 hours because of its design, not its performance: it is built to sleep. Closing the lid, low battery, throttling from poor cooling and automatic restarts after system updates each stop the strategy in the middle of the session. It is excellent as a research and charting tool, but using it as the order machine means fighting its design.

A desktop can serve as the order machine as long as all the backups in the previous section are in place and you do not run heavy backtests on it at the same time. A VPS is the least worry as an order machine. Check three things when choosing one: the data center is close to the broker's servers, it restarts itself after a disconnection and launches the order program automatically, and the remote desktop can also be reached from a phone. A VPS usually has a small processor and little RAM, so research stays on the desktop at home and the VPS holds only the strategy and the order program. That is the "desktop for research, VPS for orders" split.

You can take the next step today: on the computer you already have, run one backtest of one strategy over the longest data you plan to study, note the seconds, and multiply by the number of candidates you plan to try. That product tells you directly how many cores the backtest machine needs, and whether a graphics card is worth adding now or next year.

Frequently asked questions

Can I run program trading on a laptop?

You can use it to watch the market, write strategies and run daily-bar backtests with a small number of candidates, but it is not suited to be a 24-hour order machine. A laptop sleeps or restarts when the lid is closed, the battery is low, cooling throttles it or the system updates, and any one of these stops the strategy in the middle of the session. If you do use a laptop for orders, at least keep it plugged in, turn off every sleep setting, use a wired network, and place protective orders on the broker's server.

Can one computer handle both orders and backtesting?

Technically yes, but in practice it is not recommended. When a backtest maxes out the processor, the order program responds more slowly and extra delay appears between the close price and the order going out. A backtest machine also needs software installs and restarts often, and each of those interrupts order placement. A cheap small 4-core, 16 GB PC dedicated to orders, plus your existing computer dedicated to research, costs little and removes more than half of the risk.

What specs does a trading VPS need?

A VPS for orders has low requirements: 2 to 4 cores, 8 to 16 GB of RAM and a few tens of GB of SSD can usually run a charting program plus an order program. More important than the specs are a data center close to the broker's servers, settings for automatic restart and automatic program launch, and a remote desktop that a phone can also reach. Do not run backtests on the VPS; leave them to the desktop at home.

Does a backtest machine have to have a graphics card?

No. For a single strategy, or a daily-bar backtest of a few thousand candidates, a multi-core processor with 32 GB of RAM finishes in reasonable time. A graphics card pays off when candidates times bars becomes very large, such as one search of hundreds of millions of minute-bar candidates, where the wait goes from hours to days. Run your actual study size on the processor first, and buy the card only if the wait is truly unbearable.

References

Bailey, D. H., Borwein, J., López de Prado, M., & Zhu, Q. J. (2014). Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance. Notices of the American Mathematical Society, 61(5), 458–471.
CME Group. E-mini Nasdaq-100 Futures Contract Specs. https://www.cmegroup.com/markets/equities/nasdaq/e-mini-nasdaq-100.contractSpecs.html
Taiwan Futures Exchange. Mini-TAIEX Futures. https://www.taifex.com.tw/enl/eng2/mTX