Our Approach: Peer-Driven Accelerator Intelligence

AcceleratorAlumni predictions are different from algorithmic models. Instead of scoring startups against numerical signals, we tap into the collective judgment of people who have actually been through accelerator programs. Our approach rests on a simple premise: nobody understands what gets selected better than the people who were selected before them.

How Alumni Polls Work

Each prediction poll follows a structured process. We identify an upcoming accelerator cohort, compile publicly available information on known applicants, and present that information to our verified alumni community. Alumni are asked to rank their top 3 predicted selections based on their understanding of what the program values.

Polls run for 7-10 days to allow thoughtful responses. We target a minimum of 50 respondents per poll, though our larger polls (like the recent ICON Spark 2026 poll) attract 300+ responses. Results are tabulated using our confidence-weighted methodology and published with full transparency on sample size, response rates, and confidence intervals.

Alumni Verification

The integrity of our predictions depends on the authenticity of our community. Every member must verify their accelerator participation before gaining access to polls or community features. Our verification process is straightforward but thorough.

Applicants provide their accelerator program name, cohort year, and company name at the time of participation. We cross-reference this against publicly available cohort lists, demo day recordings, and alumni directories maintained by the programs themselves. In ambiguous cases, we request a LinkedIn profile showing the accelerator affiliation or a referral from an existing verified member.

This process rejects approximately 8% of applicants — usually people who participated in a program's free events or workshops but were not part of a formal cohort. We also periodically re-verify members whose program affiliations cannot be confirmed through public sources.

Multi-Cohort Sampling

A key design principle of our polling methodology is multi-cohort sampling. Rather than relying solely on alumni from the specific program being predicted, we draw respondents from across all 18 represented programs. This matters because different programs share overlapping selection criteria, and alumni from adjacent programs bring fresh perspectives that reduce groupthink.

For example, our ICON Spark 2026 poll included alumni from ICON Spark (weighted 1.5x), Techstars Cyber, Y Combinator, 500 Global, Alchemist, and 13 other programs. The cross-program consensus adds robustness — when alumni from very different programs converge on the same prediction, the signal is stronger.

Confidence Weighting

Not all votes are equal. Our confidence weighting system assigns higher weight to respondents whose backgrounds make their judgment more relevant to the specific prediction.

• Program relevance (1.5x): Alumni from the same program being predicted receive a 1.5x weight multiplier. They have direct experience with the selection committee's priorities.
• Recency (1.0-1.2x): More recent graduates receive a slight weight boost (1.2x for the most recent cohort) since selection criteria evolve over time.
• Historical accuracy (0.8-1.3x): Respondents who have participated in previous polls and demonstrated above-average prediction accuracy receive higher weighting in subsequent polls.
• Domain relevance (1.0-1.2x): For sector-specific predictions (e.g., cybersecurity), alumni with domain expertise receive a 1.2x boost.

The combined weight for any individual respondent ranges from 0.8x (lower historical accuracy, no program or domain relevance) to approximately 2.8x (same program, recent cohort, strong track record, domain expertise). This range is intentionally moderate — we want relevant expertise to matter without allowing individual voices to dominate the consensus.

Known Limitations

We are transparent about what our approach cannot do. Peer-driven predictions are based on publicly available information and alumni judgment — they cannot capture private dynamics like internal referrals, committee politics, or last-minute application changes. Our 72% accuracy across 45 polls means we get it wrong roughly 28% of the time.

We also acknowledge selection bias in our community. Alumni who remain engaged with the ecosystem post-graduation may have systematically different perspectives than those who have moved on. We mitigate this through broad outreach and periodic re-engagement campaigns, but the bias cannot be fully eliminated.