How we measure risk — and what we don't do

Why the DCA plan is equal-weight #

By default, your recurring-contribution plan splits the month's amount equally across your assets (equal-weight), with a fixed amount. It sounds unsophisticated. It is deliberate — and we tested it.

We tried to make our own system do something "smarter": modulate how much you contribute based on the market, and split it with our AI instead of equally. We put it through the allocation backtest (goal7/9/10/12) — six experiments with out-of-sample validation, rules frozen before we looked at the results, and a double asset universe so we couldn't fool ourselves.

• Fixed DCA beat "flexible" DCA in 7 of 7 splits: modulating the contributions leaves money idle (cash drag) and loses.
• The equal-weight split beat every allocation our system proposed (inverse-vol, momentum, band-based, full-stack) on the same book.
• The "return engine" that tried to improve the split did not generalise to the hold-out: it was overfitting, and our own validation framework caught it.

So we use the AI for what it does demonstrate value at — measuring risk — and we recommend the simple thing for the split, because that is what won. The honesty is the differentiator, not a slogan.

The Risk gauge (0–100), and why it is validated #

The Risk gauge is a number from 0 to 100 with three states — Low (<40), Medium (40–70) and Elevated (>70). It measures how much uncertainty there is right now compared with the asset's own history: when the likely range of prices widens relative to its own habit, the number rises. It is a risk reading, not a buy or a sell signal.

How it is computed (no look-ahead)

Per asset, we take today's p10–p90 band width and place it in the percentile of the asset's own prior history — "width relative to itself", using only data from before the date (zero look-ahead). The market gauge is the average of those percentiles. And it is honest by construction: an asset with no band does not score; with less than twelve months of history, it does not score; if fewer than half the assets have a band, the gauge is not shown — a number is never fabricated.

Why you can trust it: the event study

We checked whether the gauge states really precede risk, with a frozen-rule event study (goal13/14) across two asset universes and at 3- and 6-month horizons. We measured the largest drawdown that came after each state:

In all four cases, the Elevated months preceded larger drawdowns than the Medium months. This is an observed association between the level of uncertainty and the risk that followed — not a prediction of direction. The gauge also survives recalibrating the bands: its association with later risk did not depend on any calibration error. That is why it is the feature we trust most.

What we do not claim: anything about the Low state (it occurs too rarely to conclude anything) and anything that sounds like "buy" or "sell". The gauge informs your pace of saving — you can wait or split the month's contribution in two — not your direction.

The p10–p90 bands and the 80% calibration #

For each asset we show a probabilistic range of the future price: a median and a p10–p90 band — the fan that widens with the horizon. The width of that fan is the message: the more uncertainty, the wider the band. We draw no "up" or "down" arrows.

Illustrative example — not a projection for any specific asset.

"p10–p90" means that, if the world behaves like its history, the real price should fall within the band 80% of the time. And we measured it: over ten years of data the model had not seen (goal4), the real price fell within the band 80.4% of the time, measured out-of-sample. It is not a promise: it is a measured coverage.

Why we don't "tune" it further

We tried recalibrating the band and weighting the models by their accuracy. Both make it worse: recalibrating dropped the coverage from 80% to 72%, so we don't re-tune it. The disagreement between the models is the uncertainty signal — collapsing it destroys what makes the band useful. So we leave it alone.

Per-asset calibration (an honest notice)

The bands are not equally calibrated on every asset. Some assets under-cover — the range falls short more often than it should (~71–75% against the 80% target). On those assets you will see a notice next to the band. We do not correct it automatically because the recalibration does not generalise: fixing the ones that under-cover breaks the ones that are already fine. We prefer to warn rather than to correct badly.

InferenceKey's forecast engine #

The bands and the gauge come from InferenceKey's forecast engine: an ensemble of several state-of-the-art time-series forecasting models, combined and calibrated by us. The exact composition of the recipe is ours, and we keep it — but the rigor with which we measure it is entirely transparent, and it is what you can audit on this page.

What we don't do #

As important as what we do is what we rule out — almost always because we tested it and it didn't work, and our validation framework caught it. None of this is in the product:

The rule we hold ourselves to: nothing that has been ruled out comes back to the product without a new experiment with rules frozen in advance, a double universe, a hold-out and numeric thresholds. That framework already caught two mirages that looked like they were winning.