On the night of November 2, our best picks went 0 for 8. That left the record 15 units down after nine nights, and anyone checking the track record that morning would have had a clear idea of what they were looking at: a model that doesn't work. Six nights later the same record was 8.57 units up. Nothing about the model had changed.
Both of those readings are wrong, and in the same way. Fifteen nights of betting results is not much information. It feels like a lot, because each of those nights had real games and real people winning and losing. It is a good time to look at what our own numbers can and can't say so far.
What the record shows
These are the two picks per game on the site's All tab, locked five minutes before puck drop and graded after the game, at one unit each. From October 25 through November 8 there are 205 graded picks. Fifty-six won, which is 27.3%, at an average decimal price of 4.68. Total: +8.57 units, or a 4.2% return on what was staked.
Best picks, running total
Units won or lost on the All tab after each night, October 25 to November 8. One unit per pick.
Now split that same line in half. The first eight nights produced 103 picks and a loss of 7.00 units. The last seven produced 102 picks and a gain of 15.57. Two samples of about the same size, same model, same market, and they tell opposite stories. The whole profit, and then some, came on November 7 and 8, which added 20.63 units between them. Before those two nights the record was 12.06 down.
If two nights can flip the conclusion, the conclusion wasn't ready.
How much a record this size can move by luck
To put a number on it, we replayed those exact 205 picks at their exact prices 40,000 times, with each pick winning only as often as its price implies, which means a bettor with no edge at all. The bets were not made any better or worse. Only the luck changed.
Read those again slowly. A bettor with no skill and the same bets ends up at least as far ahead as we are more than one time in three. More than half of them also go through a slump as ugly as the one that made us look broken. And the honest range on our return runs from losing almost a quarter of every stake to making almost a third on it.
Part of the reason the swings are so wide is the prices. At an average of 4.68, most picks lose, and a few long winners do the rest. One 5.00 winner pays for four losing picks. That is why the line on the chart jumps around instead of drifting.
What you'd need to actually know
The standard deviation of a single pick in this record is about two units. To tell a true 4% edge apart from zero with reasonable confidence, you would need roughly 9,000 picks. At the rate the site makes them now, that is more than three full seasons of games. Anyone who tells you a tipster is proven after a good month, including a tipster showing you their own good month, is telling you something the arithmetic does not support.
The tempting cuts
This is where the real money gets lost. Open the record by market and it practically begs you to act on it:
- Assists: 2 wins from 27 picks, −18.95 units.
- Goals: 11 wins from 63, +12.95 units.
- Points: 28 wins from 72, +20.26 units.
The natural reaction is to stop backing assists and pile into points. But 27 picks is nothing. At the prices on offer, two wins or fewer from those 27 happens about 12% of the time by luck alone. It is a sign worth watching, and the model's own numbers make it a slightly louder one, since it expected more than the market did. It is not a verdict, and neither is the +20 on points.
The Multi-point tab is the same lesson at higher volume. It's a separate long-shots record: 102 picks, 9 wins, −72.79 units. It includes 38 picks on a player scoring two or more goals, at an average price of about 62, and not one of them won. That looks like a disaster. By the prices, though, you would have expected less than one win in those 38, and going zero for 38 happens about 44% of the time. A long-shot record can look terrible for a long while, and it can look brilliant for a while too, on one lucky ticket. Either way the sample says very little.
The honest part
One number in the record does deserve some caution, and it cuts against us. The model thought those 205 picks would win 31.0% of the time, and they won 27.3%. The prices implied 26.1%. So the picks did beat the market's number, but by less than the model hoped. In the replays using the model's own chances, 88% finished better than we actually have.
That is what you should expect from any system that picks its favorite bets. When you select the bets where your number and the market disagree most, you collect the real edges and your own mistakes together. So our edge is probably smaller than the model says it is, on any given night. Treat the percentage on a pick as an estimate with some air in it, not a promise.
What to do with a track record
- Count the bets before you read the result. Under a few hundred, the result mostly tells you how the bets happened to land.
- Be suspicious of the slice that looks best. The best-looking market in a small sample is usually the luckiest one.
- Look at how it got there. A profit that comes from two nights is a different thing from a profit built steadily.
- Size for the bad stretch, not the good one. A 22-unit slide in a fortnight is normal here. If that would hurt, the stakes are too big.
We publish every graded pick on the Track record tab, wins and losses both, because a record you can't check is not a record. What this one says after fifteen nights is mostly that it's too early to say. We'll keep counting.



