Impact Scores
Stakeholder impact scores estimate how positively or negatively a public company is exposed to a specific situation. They are reasoned judgments grounded in retrieved evidence — not prices, not forecasts, and not precise measurements.
In short: read the score in the official weak / strong bands. Trust direction and band. Use modest gaps (~0.15+) to rank within a response. Ignore tiny differences.
What the score is
| Property | Detail |
|---|---|
| Range | −1 to +1 |
| Sign | Positive = net beneficiary · Negative = net adversely exposed |
| Magnitude | Larger absolute values = stronger assessed exposure |
| Unit of analysis | One company × one situation (or development) |
A score of +0.70 means strongly positive fundamental exposure given the evidence — not “+70% expected return.”
How to read a single score
Impact is scored on a fixed scale used in stakeholder identification:
| Score | Meaning |
|---|---|
| −1.0 to −0.5 | Strongly negative |
| Below 0 to −0.5 | Weakly negative |
| 0 | Neutral |
| Above 0 to 0.5 | Weakly positive |
| 0.5 to 1.0 | Strongly positive |
|impact| ≥ 0.5 is the strong tier; between 0 and 0.5 is the weak tier (same on the negative side). That boundary is intentional — it is how the model is asked to distinguish material exposure from lighter exposure.
Within a band, prefer the label over the second decimal. +0.62 and +0.71 are both strongly positive; +0.18 and +0.35 are both weakly positive.
What holds up well
We re-ran stakeholder identification on the same situations with identical evidence inputs, then compared results company by company to isolate scoring variability.
| Finding | Result |
|---|---|
| Score agreement on shared companies | Correlation ≈ 0.92 |
| Typical change on a rerun | Median ≈ 0.07 |
| 95% error margin on one score | ±0.22 |
| Direction for non-trivial scores | Sign almost never flips when |impact| is clearly away from zero |
| Ranking when scores are meaningfully apart | Order stays stable |
How to read the error margin: if a company scores +0.50, a repeat assessment on the same inputs stays within about +0.28 to +0.72 roughly 95% of the time. The usual move is much smaller (median ≈ 0.07), so most retests remain in the same weak/strong band — the ±0.22 figure is the formal 95% bound, not the everyday case.
Very close head-to-heads can still swap order, the same way two analysts agree on “strong winner” but may reorder similar mid-list names.
Comparing two companies
You do not need a huge gap to rank names. A separation of about 0.15 is already a practical working threshold from retest behavior.
| Score gap |A − B| | How to read it |
|---|---|
| Below ~0.10 | Effectively tied — don’t force a rank |
| ~0.15 and above | Enough to prefer A over B |
| ~0.30 and above | Especially stable — almost never reverses on retest |
Examples
| Pair | Gap | Reading |
|---|---|---|
| +0.55 vs +0.50 | 0.05 | Both strongly positive — same tier; treat as a tie |
| +0.45 vs +0.25 | 0.20 | Prefer the first; normal, usable separation |
| +0.70 vs +0.20 | 0.50 | Strong vs weak — different tiers by design |
Neighbors in a sorted list often swap when their scores are within ~0.10. That is expected for close calls, not a reason to distrust clearer separations. Always sort by impact_score; do not rely on array order.
Practical guidance
- Read sign + weak/strong band first; then rank on gaps
- For a tighter shortlist, prefer |impact| ≥ 0.5 (strong tier); include the weak tier when you want broader coverage
- Rank when gaps are ~0.15+; collapse closer names into a shared tier
- Lean on rationales and citations for the “why,” especially on close calls
- Do not treat impact as a price target or trading signal
confidence_score uses its own bands (low → very high). Higher confidence lines up with slightly more stable retests — a soft cue, not a hard gate. Near-zero scores are the noisiest. Which companies appear can vary more than how overlapping companies are scored, especially on broad situations — design around bands and gaps rather than a frozen list order.
Short answers
Can I trust the impact score?
Yes. It is reliable for direction and weak/strong magnitude. On a single score, the 95% error margin is ±0.22 (typical retest move ≈ 0.07). It is not a high-precision measurement.
Is +0.55 meaningfully different from +0.50?
No. Same strong-positive tier; treat as a tie.
Is +0.45 meaningfully different from +0.25?
Yes for ranking — both are weakly positive, but that gap is large enough to prefer the first.
Is +0.70 meaningfully different from +0.20?
Yes — strong vs weak positive on the scoring scale itself.
If I run the same situation twice, will I get the same list?
Overlapping companies keep similar scores and usually the same band. Exact membership and very close orderings can shift slightly. Prefer tiers and clear gaps over depending on a fixed sequence.