Consider a hypothetical buyer paying $3,000 for an “HFT bot” and recording losses over six months. Neither the price nor the label establishes the strategy's operation. Investigate its actual data, execution path, costs and records before treating a retail terminal or VPS as equivalent to a low-latency institutional workflow.

What high-frequency trading actually is

HFT is a subset of automated trading that emphasises rapid processing and short-lived opportunities. Market making and arbitrage are possible strategies, but named firms need not share the same business model, holding period or hardware.

Under EU MiFID II Article 4(40), a high-frequency algorithmic trading technique combines infrastructure intended to reduce latency, system determination of individual order operations without human intervention, and high intraday rates of messages including orders, quotes or cancellations. Message counts are not trade counts. This legal definition does not require a named programming language, FPGA or a team of a specified size.

The Markets Committee's September 2011 report High-frequency trading in the foreign exchange market distinguishes HFT from broader automation and stresses the difficulty of measuring it. Its dated descriptions are not a universal present-day definition of every firm advertising high-frequency execution.

Pair the stages of each trial first

Three trial IDs each contain two consecutive, non-overlapping stages measured by one consistent clock. Calculate the median total duration.

Scroll sideways to see the full diagram

Median of three complete measurementsThree fictional trials have stage pairs 1 and 9,9 and 1,9 and 9 milliseconds. Their totals are 10,10,18 and the median is 10. Each separate column has median 9, summing to 18.STAGE MEDIANS DO NOT ADDStage 1Stage 2A10 ms1 ms + 9 msB10 ms9 ms + 1 msC18 ms9 ms + 9 msTotals:median (10,10,18) = 10 ms.Stages:median 9 + median 9 = 18 ms.

A, B, C are fictional measurements. Both stages are consecutive, non-overlapping and exhaust the measured interval. They do not describe real HFT or any broker's execution times. NIST — Measures of Location defines the median from ordered observations, selecting the middle observation for an odd count.

Block widths are proportional to time. Each pair belongs to a specific trial.

The median is not linear. Adding column medians loses which delays occurred together. Preserve trial IDs and measurement boundaries, sum each trial's stages, then sort the complete durations. Overlapping stages or unsynchronised clocks need a separate check before summing. This evaluates reporting arithmetic, not the profitability of a fast strategy.

Infrastructure differences in HFT

Colocation can place equipment close to a venue's systems. CME describes customer equipment alongside Globex matching engines in Aurora, Illinois. Check the actual connectivity and service contract; a universal annual rack price cannot be inferred from the term colocation.

Hardware, feed access, connectivity, testing and operational controls affect a low-latency workflow. Requirements vary by venue and strategy. Neither FPGA use nor a particular rebate schedule is universal, and a retail label alone does not determine access or total cost.

Measure your complete execution path from specified timestamps. Ping or a VPS's advertised latency does not measure the whole sequence through validation, routing and fill confirmation. A small hosting budget establishes no institutional execution equivalence.

How HFT affects you anyway, even if you never run an algorithm

The 2011 Markets Committee report examines HFT’s effects on participants and its behaviour in normal and stressed markets. Its dated analysis does not guarantee a particular retail account’s spread or assign all price improvement to HFT. See market makers. The Markets Committee’s January 2017 release on the 7 October 2016 sterling flash event identifies several factors, including the time of day, option hedging, stop orders and algorithm use. Reducing that event to “HFT switched off” would discard that evidence.

Liquidity and execution can deteriorate during stress, but mechanisms and fills need evidence for each episode. For a hypothetical ordinary stop at 1.18, a later fill at 0.90 illustrates slippage only; it is not a documented SNB-event trade. Check the contract's stop type and execution terms instead of assuming a fixed price remains available.

What you can realistically do instead of chasing HFT

In a hypothetical model, each signal wins 0.5 pip with probability 55% and loses 0.5 pip otherwise. Gross expectancy is 0.55 × 0.5 − 0.45 × 0.5 = 0.05 pip. Total entry-and-exit costs of 0.10 pip change this to −0.05 pip; costs of 0.50 pip change it to −0.45 pip. If a simulation already uses executable Ask entry and Bid exit, spread is already included: do not subtract it twice. Real probabilities, payouts and costs require measurement.

Fill rate alone can conceal a different result. In a controlled hypothetical exercise, freeze ten opportunities: five would return +1 pip and five −1 pip, before costs. Execution A fills three of each, returning zero from six fills; execution B fills one positive and five negative, returning −4 pips, or −0.67 pip per fill. Both fill 60% of opportunities. These assigned paths isolate selection; real unfilled orders do not have a known counterfactual execution price. Record rejections and the characteristics of filled opportunities as well as speed.

Choose the strategy's observation and execution requirements before choosing hardware. Scalping or an M5–H1 algorithm still faces spread, slippage and timing risk; a longer bar does not make latency irrelevant or establish profitability. See first steps in algorithmic trading.

Red flags — the "retail HFT" course as a business model

For an HFT-branded product, request reproducible rules, complete results and execution assumptions. Investigate backtest selection, unavailable future information and conflicts of interest. A marketing label alone proves neither legitimacy nor fraud.

The CFTC’s 25 January 2024 customer advisory warns that high or guaranteed returns claimed for AI bots and signal services are fraud red flags. A price tag or closed source code alone is not proof of fraud. Ask which data, costs, omitted orders and evaluation period support the result before comparing it with an institutional workflow.

Check the relevant regulator's register and warnings for the exact entity and activity. For UK services, the FCA Firm Checker helps verify authorisation and permissions. A software product or an algorithm label does not establish the legal status of every service sold with it.

Your next step — how to stop wasting time on a false promise

  1. Inventory the product’s evidence. Record its purchase and subscription costs, actual live results and operating period. Separate a demo, a backtest and the live account; a marketing name or purchase price cannot establish how the system works.
  2. Verify the exact entity and service. Use the relevant jurisdiction's official register and warning material. For UK services, check FCA permissions and registered contact details. Neither absence from a warning list nor a matching brand name guarantees legitimacy; recovery depends on the actual facts and protections.
  3. Define a small research project. Specify one strategy, data set and cost model, preserve an untouched evaluation window and test order handling before considering deployment. Six months is a possible study budget, not evidence of readiness.
  4. Evaluate the complete workflow. Review exposure, costs, execution and records against the actual requirements. M30 or a particular broker execution model does not guarantee an advantage after a year.