baseedge_labs
← research

Is Tokyo Drift's 2024 losing streak a curve fit or a different market? Neither, it turns out

Sep 3, 2026 · BaseEdge Labs research

Tokyo Drift

Tokyo Drift’s MGC configuration was hand-tuned during 2026 and forward-tests well. Backtested over the full sample from January 2024 to August 2026, its edge is thin. Two readings of that gap are possible and they have opposite implications:

They are indistinguishable if you only ever test the tuned config. So three earlier default sets were run on the same bars, all frozen before 2025-26 existed and none ever tuned on this window.

Method

What we found

  1. The tuning is not why 2024 loses. A config that never saw 2025-26 lost 2024 just as hard as the one tuned on 2026. Quarter by quarter, the Default and MGC-tuned results moved together almost exactly.
  2. The control confirms it. v1.0.0, a different entry model on identical bars, was anti-correlated with the Default. So the close match between Default and MGC-tuned is not just two configs sharing a market. The quarter drives the result, not the settings.
  3. The market did change. Over the sample, gold’s price rose steeply and its average daily range and ATR rose much faster. 2024 was a far quieter market than 2026.
  4. Volatility is not the explanation, though it first looked like one. Split by ATR as a share of price, the high-volatility quarters appeared clearly net positive and the low-volatility quarters lost. Take out the two best quarters and the effect disappears. The remaining high-volatility quarters were statistically indistinguishable from zero, and the tuned config did worse in them than in the quiet quarters.
  5. There is no dose-response and no volatility floor. 2026Q1 was the most volatile quarter in the sample. It made only a little on the Default and lost on the tuned config. Within the high-volatility quarters, more volatility went with lower returns for both configs.
  6. The whole profit is two quarters. Over the 2.5 years, both configurations made all of their net profit in 2026Q2 and 2026Q3. Most of the other quarters lost money, and taken together the rest of the sample was a net loss.
  7. Those two quarters do not reduce to a short bias. They were gold breaking down from a parabolic top, followed by a very wide-range quarter. Longs and shorts both paid in both. No regime variable tested here separates them from the rest of the sample. The honest description is that the edge is episodic and unexplained. No condition for it has been found.
  8. The untuned defaults are the better configuration. Compared with MGC-tuned, the Default had a higher win rate, lower maximum drawdown and more net-positive quarters in all four windows. It also had the higher profit factor and net result on the full sample, from April 2025 and over the last 12 months.
  9. The tuned config leads only in the last six months, on profit factor and net, at a lower win rate and with a deeper drawdown. That is the signature of a configuration fitted to a recent window rather than a better one. It is a separate question from whether 2024 was a fit artifact, and it lands the other way.
  10. v1.0.0 did not share in 2026. It made money in some quarters where the others lost, missed the 2026 run, and finished the full sample slightly negative.
  11. Every session was net positive on the Default. Run alone over the full sample, Asia, London and NY each finished net positive on the Default config, with London the weakest. Under the MGC-tuned config, London loses money. Session results run alone do not add up to the combined run, because trade budgets and reversal slots interact when sessions run together.

One deliberate deviation from “untuned”

The published Default is not quite the raw pre-release values. It ships Direction Gate = Slope only rather than Side + Slope.

The gate is the one change out of eleven research stages and roughly 1,300 configurations that survived a walk-forward. It defaulted to the old value in v1.4.0 and v1.5.0 only so that each version would behave exactly like its predecessor at default settings. That was a compatibility choice, not a recommendation. Those raw values are a backward-compatibility baseline, not a curated config. Shipping them unexamined would have published a default that ignores the only validated finding the research produced.

On the full sample, the published Slope only Default matched the raw Side + Slope values on profit factor. It finished slightly ahead on net with a lower maximum drawdown.

Several of the tuned parameters are inert in both configs, incidentally. erPeriod, erThreshold, momThreshold, maxExtensionStdDevMult and vwapDistAtrMult all sit on filters switched off in each. The effective difference between the two configurations is five settings, not twenty-two.

What this does not say

This is one instrument, one timeframe and 2.5 years. These are hypothetical backtests, not broker fills. Nothing here tests MES, MNQ or MBT, where the same logic may sit in a different relationship to the market. And an edge concentrated in two quarters is a small enough base that the next one is genuinely uncertain in both directions. This data cannot distinguish an edge that arrived in 2026 from an episode that has already passed.