The machine learning is performed via tuned XGBoost algorithm, maintaining few models for classification of score sum, score difference, both teams to score and correct score.

Our models are validated via bootstrapped walk-forward routine where test data is the last 365 days, while train data is all earlier.

Our models are trained every day, using results of previous game day - focusing rather on proper features than on exact weights and biases.

Calculated classification probabilities are adjusted then via Dirichlet-multinomial game score distribution, and for case of Asian half- and quarter- odds respective quasi-probabilities are calculated.

While generally producing balanced opinion on game results, betting market is affected by money flow and preventing arbitrage. These and other factors, combined with stochastics of the whole process, give our algorithms sweet spots where probability of the event calculated by us is greater than what bookmakers state, even including their margin.

Comparing calculated and market probabilities, we receive optimal event to bet on.

Our historical results show that:

• 94% of all games contain at least one bet with positive expected long-term yield

• 11% of all bets in all games have positive expected yield

• 0.05% of all bets (800 bets a year) promise expected yield of 10% and greater

• 66% of all games contain at least one bet with expected robust wager of at least 1% of the budget

Having probabilities calculated and odds collected for the game, we select optimal prediction - a tip where the bet could be placed at.

For optimality criterion we elaborated the simple yet efficient approach, minimizing inevitable temporary losses.

This algorithm is used when we publish predictions at our website or make a tip at 3rd party portals.

To calculate an optimal wager for a bet, a mathematical technique aiming profit maximization is combined with another technique aiming risk restriction.

For profit maximization we use results of Kelly.

For risk restriction we use own results coming from research of Random Walk process.

Our benchmark tests show that even in perfect case where all probabilities are calculated without error, the pure Kelly method ruins in 20% of cases. Moreover risk grows significantly, as inevitable mistakes take part in probability calculations.

Applying our risk restriction technique just slightly reduces resulting profits, while increases robustness a lot - even in case of slightly optimistically calculated probabilities this approach leads only to 1% of ruin, and negative effect of increased mistakes is less comparing to pure Kelly case.

Having proper risk management, it is possible to select optimal wager for arbitrary betting odds received from any bookmaker, or your neighbor in the bar.

To make betting even safer, we add artificial skepticism to our bets via discounting calculated probabilities a tiny bit.

One more important practical thing about our bets is that we play on original fortune, without necessity to add new money to it.

Having plenty of betting odds, we build optimal betting strategies for bookmakers automatically, while also apply on manual betting via different platforms, thus providing wide range of opportunities for our users.

Betting opportunities are not restricted with supported bookmakers and platforms - quite contrary, it is possible to get and optimal wager, just knowing betting balance and odds of the event.

Consider not betting, if we say 0.

Consider also number of bets: rough calculations recommend making not more than 10 bets a day, to stay relatively safe.

Keep in mind, there is no way to eliminate the risk of losing; at least mathematics says so - but our mathematics is to reduce it to acceptable margin.

We are absolutely transparent regarding history of our predictions and bets - giving you a true sight on where we win and where we lose.