Trixter MCD

The robust covariance (mincovdet) method, or Minimum Covariance Determinant (MCD), is a robust method for estimating the covariance matrix and mean vector that is resistant to outliers in the data.

The idea of the method

Classical sample covariance is very sensitive to "contamination." MCD solves this problem by finding a "clean" subset of observations that does not contain outliers and computing standard estimates (mean and covariance) only on it.

What MCD (mincovdet) is used for

1. Detection of multivariate outliers
This is the main application. MCD provides robust estimates of the mean and covariance, from which Mahalanobis distances are computed. Observations with a large distance are flagged as outliers. Unlike the classical approach, the estimates themselves do not "drift" because of outliers, so they are seen correctly.

2. Robust estimation of the covariance matrix
When data are contaminated, ordinary sample covariance is heavily distorted. MCD provides a robust alternative — for example, for:

• portfolio optimization in finance (robust risk estimates),

• constructing confidence regions,

• any tasks that require an honest covariance matrix.

3. Robust estimation of the distribution center
The mean vector computed from the "clean" subset is also resistant to outliers — unlike the ordinary mean.

4. Preprocessing before other methods
MCD is often used as a data cleaning step: find outliers → remove/reduce their weight → then apply ordinary methods (regression, clustering, PCA) that are themselves unstable to outliers.

The Trixter MCD bot is built using this approach.

The Trixter MCD algorithm implements pattern discovery using the Minimum Covariance Determinant (MCD for short). We proceed from the assumption that the gold market is sufficiently efficient and that most price fluctuations form market noise that cannot be predicted. But we can identify fragments of the gold chart (so-called anomalies) in which it is maximally predictable. Thus, I identified 40% of chart segments that are well predictable.

Key principles

• Analysis of gold volatility across multiple timeframes and indicator periods.
• Grouping volatility to find effective patterns (from 100 to 3000).
• Statistical validation: high mathematical expectation of profit and statistical significance.
• Patterns that fail validation are marked as noise — in such situations, trades are not opened.
• A mandatory 2-year forward test before being allowed on a real account.

Features of the trading logic

• Opening trades with limit orders at a selected distance from the price — minimal slippage.
• If a limit order is not triggered (the price did not touch the level) and the signal disappears — the order is deleted.
• Open positions are accompanied by stop-loss and take-profit, and are closed at market on an opposite signal.
• The robot does not use martingale, arbitrage, or other high-risk strategies.
• Recommended timeframes: from M5 to H1. Optimal — H1 XAUUSD.
• Flexible configuration of trading activity is provided (parameters below).

Main settings

• Allow BUY signals, Allow SELL signals — allow or prohibit buys/sells when there is confidence in the market direction.
• Signals sensitivity, Filter sensitivity — threshold for signal sensitivity and noise filtering. The default value is 0.5. It can be increased to 0.7–0.9 for cleaner signals.
• Filter by trading hours — restriction of trading time. By default — around the clock.

Money management settings

• Distance in points for limit order — distance from the price to the limit order. Default: 250. Smaller → more trades, larger → more conservative.
• Max positions + orders number — maximum number of simultaneously open positions (without averaging). Default: 3.
• Additionally configurable: delay between openings, maximum spread control, stop-loss and take-profit levels.