Why most backtests lie
A backtest means testing your strategy on old games to see if it would have made money. It sounds simple, but it's shockingly easy to fool yourself. This lesson shows the two most common traps and how VEKTORIUM avoids them.
Trap 1: accidentally using the future
Say you're testing whether you could have predicted April's games. If your model uses stats from the WHOLE season — including May and June — then April's 'predictions' already know the future. That's called look-ahead bias, and it makes any backtest look amazing and be worthless. VEKTORIUM's trainer only uses stats from before each game's date, so nothing 'sees' the future.
Trap 2: too few games
With only 30 games, even a coin flip can look like a great strategy if it gets lucky. Metrics need hundreds of games before they mean anything. A backtest over 30 games tells you nothing; over 500 it starts to be useful, and even then, results come with uncertainty.
How honest testing works
The gold standard is called walk-forward testing: train your model only on the past, then see how it does predicting the next games it hasn't seen, then move forward and repeat. Every prediction is made with only the data available at that time — the same situation you're in for today's real games.