# freqtrade hyperopt: matching results, 6–11× faster in our published tests — would you want it on your own machine? Waiting for hyperopt can slow down strategy iteration. We measured complete runs on a public fixture and checked whether the reported results stayed the same. ## What we measured We checked fifteen cases: freqtrade's own SampleStrategy and nine strategies from the public freqtrade-strategies repository, on 5-minute, 15-minute, 1-hour and 4-hour candles, with 263 epochs and 204,132 trades in total. Every compared result value matched: each epoch's parameters, loss and metrics, and every trade with every order. The comparison ignores two wall-clock timestamps per epoch. One case uses freqtrade's own engine as a fallback; the linked table identifies it. The timed runs below used SampleStrategy. They are not speed measurements for all fifteen cases. The factors are for the elapsed time of the whole process, on one CPU core. | Run | Zen 4 (one CPU core) | Zen 3 (one CPU core) | |---|---|---| | 30 epochs, 7 interleaved rounds | 5.91× | 6.72× | | 100 epochs, 2 interleaved rounds | 7.18× | 11.02× | In these measurements, the 100-epoch runs had larger factors than the 30-epoch runs. The full method, hardware, limits and per-case table are on the [showcase page](/showcase/freqtrade-hyperopt/README.html). ## Why you cannot download it yet The published study uses freqtrade 2026.9, licensed under GPL-3.0, with our build loaded into the same process. We have chosen not to distribute that research build; today it is available through hosted runs. A local product would need an appropriate implementation and distribution model. For traders who want to keep their strategy and data on their own machine, we are considering the local option below. ## What we are considering A separate program that runs the hyperopt of your freqtrade strategy on your machine: Windows, macOS and Linux, a one-time purchase, your strategy and data stay where they are, and the result file is checked against freqtrade's own output. It is not written yet; the measured factors above are for the current version, and we will measure the local one before we promise any number for it. Before we commit to building it, we want to know whether it matters to you, and what the saved time would go to. If a faster local hyperopt would change how you work, [tell us here](/?src=ft-local#start): a name and an e-mail are enough, no payment and no obligation. In the free-text field, one line helps us most: - what you would do with the time: more epochs, more pairs or timeframes, walk-forward or out-of-sample checks, or nothing in particular; - your operating system and whether you use FreqAI. If fewer people answer than we need, we will say so on this page.