UNDERSTANDING PICKS
Why a high win probability does not guarantee betting value
Win probability asks how often an outcome might happen. Betting value also depends on the price. A team can be more likely to win than its opponent and still be priced above a defensible estimate of its chance.
That distinction matters when reading a list called “top picks.” A list can rank likely winners, estimated value, or some other measure. Monitor Sports Picks currently ranks estimated winner probability. It has not demonstrated a profitable edge or a validated historical win rate.
A 70% winner at −300: the arithmetic
Consider a hypothetical team estimated to win 70% of the time at a price of −300. A 300-unit stake earns 100 units of profit when successful and loses 300 when unsuccessful. The price needs a 75% success rate to break even: 300 ÷ (300 + 100).
If the 70% estimate were correct, the expected net result per identical 300-unit stake would be:
This is an average over the assumed probability distribution, not a forecast of the next game's result. A single game yields either outcome; it does not return the average. The estimate can also be wrong.
Winning most selections can still produce a loss
Imagine ten hypothetical selections, each at −300 with a 300-unit stake. Seven win and three lose. Seven profits of 100 total 700; three losses of 300 total 900. Net result: −200 units despite a 70% win rate. Total amount staked is 3,000 units, so net return divided by amount staked is about −6.67%.
Those ten results are an illustration, not this site's record. Changing the prices changes the conclusion. A reported win percentage without the actual odds and staking assumptions leaves out information needed to evaluate financial performance.
A simple framework for comparing claims
| Measure | What it tells you | What it does not establish |
|---|---|---|
| Win rate | Share of settled picks that won | Profit without prices and stakes |
| Net return | Profit or loss under stated assumptions | A repeatable future advantage |
| Calibration | Whether probability groups align with observed outcomes | A favorable price on every selection |
What a useful prediction record should retain
A record should preserve every published selection, its publication time, available price, probability estimate, source, and eventual outcome. It should state how voids, pushes and cancellations are handled. Keeping losing selections matters as much as keeping winners. Without this, selective reporting can make a weak method appear stronger.
Calibration asks a different question from win rate. Across a sufficiently informative sample of picks labelled around 70%, did roughly 70% win? A few wins or losses cannot establish that. Changes to the method should be documented so an older result set is not silently presented as validation of a new model.
Same win rate, different uncertainty
Seven wins from ten selections and 70 wins from 100 both produce a 70% observed win rate. They do not provide equally precise evidence. Using the Wilson interval described by NIST, the hypothetical comparison is:
| Hypothetical record | Observed win rate | Interval for a common win probability |
|---|---|---|
| 7 wins / 10 selections | 70% | 39.7%–89.2% |
| 70 wins / 100 selections | 70% | 60.4%–78.1% |
This illustration assumes independent binary outcomes with a stable common success probability. Sports selections can have different probabilities, share teams or conditions, and change with the model. Those differences limit a simple binomial analysis. The intervals are not next-game forecasts or profit estimates, and do not validate this site's predictions. A larger sample also cannot repair omitted losses or a retrospectively chosen reporting period.
Check calibration against the actual forecasts
The Met Office's explanation of reliability diagrams compares forecast probability groups with observed event frequencies. The same statistical idea can help assess sports forecasts; the source discusses weather, not this site's model.
In an original hypothetical group of 100 selections, suppose the average published probability is 68% and 60 win. The observed rate is 60%, eight percentage points below the group's average forecast. Compare against 68%, not a rounded “70%” label. Show the number of selections in every group, use groups chosen before inspecting results, and repeat the comparison on future data. A single gap does not by itself establish a persistent bias.
Compare probabilities with a retained market baseline
A model and a market baseline can select the same winners while assigning different probabilities. To check whether the model adds information, retain both forecasts for the same team, market and observation time. Comparing different games or a pregame model with later prices answers a different question.
One probability-error measure is the binary Brier score: average (p − y)2, where p is a probability between 0 and 1 and y is 1 when the defined event occurs and 0 otherwise. Lower is better on the same evaluated sample. Beth Ebert's probability-verification workshop, hosted by ECMWF, explains this score and comparison with a reference forecast. That is a weather-methods reference, not validation of a sports model.
Here is an original fictional three-game example. Each row concerns a named team's win in a completed two-outcome game. The baseline probabilities are invented normalized market inputs, not actual bookmaker quotes.
| Game | Model p | Baseline p | Outcome y |
|---|---|---|---|
| A | 0.80 | 0.70 | 1 |
| B | 0.70 | 0.60 | 0 |
| C | 0.60 | 0.55 | 1 |
Both methods favor the listed team in all three games, so both choose two winners out of three. Yet the model's score is (0.04 + 0.49 + 0.16) / 3 = 0.2300, while the baseline's is (0.09 + 0.36 + 0.2025) / 3 = 0.2175. The baseline has lower error in this invented sample; winner-count accuracy alone hides that difference.
Three games cannot establish a lasting advantage. Preserve an evaluation rule before results arrive, report sample size and missing cases, and handle draws or unresolved games explicitly rather than silently coding them as losses. This binary example does not evaluate a three-outcome market. A lower probability-error score also does not establish profit at available prices.
For our experimental blend, this comparison would require retained model estimates and the contemporaneous normalized market for the same selections. We do not claim such a validated performance advantage. See how to preserve the price snapshot and what our calculation actually uses. Source and arithmetic checked October 8, 2026.
A practical checklist for auditing a picks record
- Fix the reporting window. Ask for every selection in a stated period, including losing days, rather than a screenshot of the best streak.
- Check the original entry. Retain the pregame timestamp, selection, market rules, price and probability. A later edited percentage is a different forecast.
- Reconcile the totals. List wins, losses, pending games, pushes and voids separately. State which outcomes enter each denominator. Check the market's draw and time-period rules before classifying a result.
- Separate versions and assumptions. Identify model changes and distinguish actual stakes from simulated returns. A backtest used to choose the model is not a prospective test of that choice.
- Ask what remains unmeasured. Missing prices prevent a return calculation; missing original probabilities prevent a calibration check. Neither gap is filled by a high win percentage.
Statistical references checked September 19, 2026. The record comparisons and audit checklist are educational examples, not an audited performance report.
What is the chance all five picks win?
Individual win probabilities do not describe the chance of a clean sweep. In an original hypothetical example, suppose five selections each have a true 70% chance of winning and their outcomes are mutually independent. The chance that every selection wins is:
Under those assumptions, the chance of at least one loss is 1 − 0.16807 = 83.19%. That is not the chance all five lose: all five losing would be 0.305 ≈ 0.24%. Each selection remains more likely to win than lose, while a perfect five is much less likely.
Penn State's lesson on independent events explains the multiplication rule and mutual independence. Multiplying standalone probabilities requires that assumption; putting selections in different rows does not establish it. For two dependent events, use P(A and B) = P(A) × P(B given A), rather than substituting the standalone probability of B.
For a deliberately extreme counterexample, imagine five identical copies of one 70% event. All five happen together, so their joint probability is 70%, not 16.81%. This is a mathematical illustration of dependence, not a description of our selections or a realistic sports forecast.
Our board does not estimate joint outcomes or validate independence. Its individual estimates also remain experimental. Do not report the product of the displayed percentages as a tested chance of a daily sweep. Keep winning outright and covering a spread separate when defining which events you are counting.
Probability reference checked September 25, 2026. These calculations describe stated assumptions, not this site's record or today's games.
What the daily five means today
Our board uses normalized odds and, where available, a limited record adjustment. It does not independently model injuries, lineups, starting pitchers or every matchup factor. Heavy favorites can rise to the top because the list prioritizes estimated likelihood of winning. The weights are experimental choices, not fitted proof of an edge.
Use the public methodology to understand those limits. For the underlying price arithmetic, see moneyline odds and implied probability. A confident-looking percentage is only as useful as the evidence behind it.
All examples and calculations are hypothetical and original. Moneyline payout conventions can be checked in Caesars: Sports Wagering Basics (PDF). No historical performance is claimed. Source checked September 17, 2026.