Research
Signal attribution
Why the model favours one name over another today, in the units of the forecast: probability, basis points or expected return.
Solutions · Capital markets
Attribute every forecast to its inputs, see which regime the model learned in, and catch decay while it is still a line on a chart. For quantitative research, risk and model validation teams.
Where it fits
Any model that makes one forecast from a fixed set of features: a probability, a return, a volatility, a cost.
Research
Why the model favours one name over another today, in the units of the forecast: probability, basis points or expected return.
Risk
Volatility, liquidity, prepayment and credit-spread models are numeric predictions. Kaimei explains each one in its own units and tracks it for drift.
Validation
Behaviour curves show whether a model responds to an input the way theory says it should, and whether that response is monotonic.
Production
Every input compared with the training distribution on a schedule. Attach realised outcomes and track hit rate or error by day, week or month.
Release
Ship a retrain as a new version and compare its insights with the one it replaces before it takes risk.
Execution
Slippage and fill models built on fixed features, explained fill by fill.
A worked example
Public data49 Industry Portfolios, Kenneth R. French Data Library (Fama & French).sector-rotationscikit-learn · gradient boosting
A gradient-boosted model estimates the chance that an industry beats the median industry next month, from nine inputs known at month end: momentum, reversal, volatility, beta and three market-wide measures. It is trained on July 1963 to December 2009 and scored on every month since. A research example, gross of costs, and not investment advice.
POST https://api.kaimei.ai/v1/explanations
Outcomes
Kaimei does not judge a strategy. It shows what the model does, and the picture here is plain: an edge of 5.2 points in sample that is close to nothing afterwards.
The attribution says why. In August 2026 the model's top pick is electronic equipment at 56.3%, and +5.9 of the 7.2 points above average come from a single input: its momentum rank. A model that is mostly one well-known factor inherits that factor's fate.
What it learned
momentum_rank is the most influential input, followed by reversal_1m and market_volatility_12m. Together they carry 68% of the measured importance (permutation importance over 2000 rows).
3.7% of decisions flip under 5%-σ input noise. Most fragile inputs: momentum_rank, volatility_12m. Kaimei rates that medium severity and raises it as an alert: On 2,000 reference rows the model predicts outperform 42% of the time; 99% of cases sit within ±0.10 of the decision threshold.
momentum_rank increases the probability of outperform, most sharply between 0.8425 and 0.9309 (range 46% → 54%).reversal_1m increases the probability of outperform, most sharply between -3.314 % and -1.709 % (range 47% → 50%).momentum_rank is the most influential input, followed by reversal_1m and market_volatility_12m. Together they carry 68% of the measured importance (permutation importance over 2000 rows).Regime monitoring
Each five-year window of live forecasts is compared with the 1963 to 2009 training rows. By 2015 to 2019 the market-wide inputs had already left the training distribution, and 2020 to 2024 adds the industries' own volatility. Stock-level inputs such as momentum rank stay stable throughout: the signal looks the same, and the conditions around it do not.
market_return_12m, market_volatility_12m, dispersion_1m, momentum_12_1. 2,940 logged predictions against 5,000 reference rows.dispersion_1m, market_volatility_12m, volatility_12m, market_return_12m. 2,940 logged predictions against 5,000 reference rows.Population stability index per input. Bars are cut off at 0.5; the line marks 0.25, where Kaimei calls drift significant.
One forecast
At the end of March 2020 the model ranks aircraft last of 49, at 43.9%. The reasons are on the record: a −35.8% month (−2.5 pts), weak momentum (−2.6 pts) and a beta of 1.61 (−1.6 pts).
The industry returned +7.0% the next month against a median of +12.8%. Any logged forecast can be explained this way later, by its prediction id.
Who it is for
See which inputs a model really uses before it is sized. Compare two versions of a signal by what they learned, not only by their backtests.
Check direction and monotonicity against theory, measure fragility under input noise, and keep the evidence with the model's record.
Ask why for any position the model drives, in the dashboard, and see an alert when the inputs leave the range the model was trained in.
Rules and guidance
| Rule or guidance | What it asks | Where Kaimei helps |
|---|---|---|
| Algorithmic trading in the EUMiFID II, Commission Delegated Regulation (EU) 2017/589 (RTS 6) | Testing of algorithms before deployment, real-time monitoring and an annual self-assessment. | A stored analysis at each release, drift checks on a schedule, and a record of every forecast and its explanation. |
| Algorithmic trading in the UKPRA Supervisory Statement SS5/18 | Governance, approval, testing and an inventory of algorithms. | Versioned models with their insights, exportable for the inventory. |
| Model risk guidanceUS interagency guidance of April 2026 (SR 26-2); PRA SS1/23 in the UK | Validation and ongoing monitoring of pricing, risk and valuation models. | Behaviour curves, sensitivity and stability evidence, and monitoring with alerts. |
Kaimei produces evidence for these processes. It does not make a model, a notice or a filing compliant, and nothing on this page is legal advice. Rules are cited as they stood in November 2026; check what applies to you now.
Coming soon
What we are building now, and what to use until each one ships.
Models that read ticks or a price history step by step, with each prediction attributed to time steps as well as inputs. Until then, a model on lags, rolling statistics and other fixed features works.
Every horizon or target of one model explained in a single call. Until then, register one model per forecast horizon or target.
An alert when what the model relies on shifts, not only its inputs: a signal that stops mattering, or one that starts to carry the model.
A simple, interpretable model fitted to the same reference data, so validators can see what the complex model adds. Until then, register a challenger yourself and compare the two side by side.
Said plainly. Kaimei explains and monitors the models you bring. It does not establish cause and effect, and nothing it reports is investment advice.
What works today, and everything on the way, is in Supported models and what’s coming.
Bring a signal or a risk model. We will show you what it is leaning on.