Recommended Free Tools
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
In Michael Hairetis’s five-session Nasdaq-100 futures evaluation, none of four ways of applying TimesFM 3.0 beat the article’s naive comparison. The reported outcomes vary by target and metric, however: forecast errors, directional accuracy, agreement with another method, and a detector’s hit rate are not interchangeable measures.
What the test covered
Hairetis’s DEV Community article, posted October 1, 2026, describes 118,857 Nasdaq-100 futures bars sampled every three seconds across five sessions. The data included bid/ask volume. The targets were midpoint price and an order-flow series described as cumulative delta minus volume signed by candle direction.
The author says the evaluation was strictly causal and confined to sessions. The accessible article text does not provide enough implementation detail to independently assess those safeguards. The reported results are a single author’s experiment, not an independently replicated benchmark.
Four approaches and their reported results
| Approach | Reported result |
|---|---|
| Direct forecasts on three-second bars | Across 2,674 anchors with a 36-bar horizon, price error was 7.4% worse than naive, with 50.0% directional accuracy; order-flow error was 7.4% worse, with 54.4% directional accuracy. [DEV Community article] |
| One-minute aggregation | At a 36-minute horizon, price error was 5.4% worse than naive, with 50.2% directional accuracy; order-flow error was 19.2% worse, with 44.9% directional accuracy. [DEV Community article] |
| Anomaly-style residual | The author summed rolling z-scores for the paired series to form a mean-reverting residual. Its error was reported as 5.6% worse than naive, with 51.6% directional accuracy. [DEV Community article] |
| Quantile-derived detector | Using the nine forecast deciles to estimate up probabilities for each series and combine them, the detector had +0.000 agreement with the incumbent method. Its highest-confidence group had a 24.7% hit rate, below the 34.3% base rate. These are agreement and conditional hit-rate figures, not forecast-error scores. [DEV Community article] |
The article’s indexed text does not establish the precise naive baseline implementation, the sessions’ dates, the market-data vendor, transaction costs, slippage, confidence-bucket construction, uncertainty intervals, or reproducibility artifacts. Those omissions matter when interpreting small differences or considering live use. Hairetis also says an apparent exception proved to be a sampling artifact and refers to a follow-up post; that follow-up is not verified here.
#1 Best Overall
How to interpret the comparison
“Worse than naive” is specific to the reported error measure
For the direct forecasts, one-minute aggregation, and anomaly-style residual, the author reports error relative to a naive comparison. A percentage worse on that error measure does not mean the same thing as directional accuracy. For example, the one-minute price forecast was reported as 5.4% worse than naive while its direction was right 50.2% of the time. Neither figure alone establishes whether a strategy could make money.
The detector result answers a different question
Zero reported agreement with the incumbent method indicates that the two methods did not align under the author’s reported comparison; it is not an error score. Likewise, a 24.7% hit rate in the highest-confidence group is a conditional result, compared with the stated 34.3% base rate. The available details do not define the group construction or uncertainty around those rates, so they should not be treated as a general estimate of future detector performance.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the results do—and do not—establish
Within this five-session sample, the article reports no improvement over its naive comparison across the forecasting and residual framings, and the quantile-derived detector did not outperform the stated base rate in its highest-confidence group. Aggregating to one-minute bars did not rescue the reported errors, although the price and order-flow outcomes differed.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →This experiment does not establish that TimesFM 3.0 fails on other instruments, periods, market regimes, or deployment setups. Nor does it establish profitability or account for trading costs. It is best read as a caution against assuming that a capable general time-series model, a coarser sampling interval, or a more tailored framing will automatically produce a useful futures signal. The underlying article is Michael Hairetis’s DEV Community report; the figures above are attributed to that report rather than independently validated.
Quick Recap
Best Value
Rank #4
- Used Book in Good Condition
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
