# Pre-registration — Daall Index feasibility tests Registered: 2026-10-29 (committed before any test result was computed; the git timestamp is the record). Data: Etimad award records (read-only extract of all awards since 2019 from the collector's database), Tadawul issuer list, quarterly revenue per issuer (stockanalysis.com, cross-checked against Yahoo Finance), daily prices (Yahoo Finance), Ministry of Finance quarterly budget performance reports. Rules that apply to every test - Point in time: an award counts from the day it could have been known. From 30 Jul 2026 that is the first day our collector saw it (`tender_change.is_awarded`); before that, award date + 30 days (the measured 90th percentile of the publication lag; median 3 days). - Values: award values include 15% VAT; divided by 1.15 when compared with revenue or market value. - Ticker mapping: registered (legal) Arabic name containment only (precision 20/20, recall 20/21 on the largest winners); trading-name matches are not used. - Every result is published, pass or fail. No test is added, dropped or changed after results are seen; any later variant is labelled exploratory. ## T1 — Contracted backlog predicts listed issuers' revenue (firm level) Expected government revenue of issuer i in quarter q = sum over its known awards of value/1.15 × (days of the contract falling in q / contract days); contracts start on the award date; contracts without a stated duration take the median duration of their award type and are counted separately. Metric: Spearman correlation across firm-quarters between ΔE = (E[q] − E[q−4]) / R[q−4] and ΔR = (R[q] − R[q−4]) / R[q−4], with R reported revenue. Inference: firm-clustered bootstrap (resample issuers, 2,000 draws). Pass: correlation > 0 and the 95% bootstrap interval excludes 0. Secondary (reported, not decisive): issuers whose expected government revenue is at least 20% of their trailing revenue. ## T2 — Awards large relative to the issuer move its share price Events: awards to mapped issuers with value/1.15 at least 1% of the issuer's market value at the previous quarter end (market value = price-to-sales × trailing-four-quarter revenue). Event day: first trading day on or after the visibility date defined above. Return: issuer return minus TASI return, compounded over days 0 to +10. Pass: mean abnormal return of these events > 0 with the 95% interval (bootstrap over event dates) excluding 0. Reported alongside: events below 0.1% of market value as a placebo. ## T3 — Award value leads government spending (macro) Quarterly award value (all awards, VAT included, no cap) against MoF quarterly expenditure on use of goods and services plus non-financial assets, both as year-on-year growth. Metric: Pearson correlation of award growth at lag k = 1..4 quarters with spending growth. Pass: at least one lag with correlation > 0 and p < 0.0125 (0.05 Bonferroni-corrected for four lags). With about 20 quarterly observations the test has low power; a fail is reported as "not shown", not as "no relation". ## T4 — Coverage of the award record (descriptive) Annual award value against annual MoF expenditure on goods and services plus non-financial assets. Reported: the ratio per year and its coefficient of variation. No pass/fail. --- # Second registration — T2′, T2″, T5 Registered: 2026-10-30, before any of the data these tests use was computed; the commit that adds this section is the record, and its hash is published on the evidence page. The general rules above apply unless stated otherwise. These tests answer the independent review of 30 Oct 2026: T2 used a conservative visibility rule (award date + 30 days) and its placebo was significant, so a fail there may reflect the return measure and not the absence of an effect; and the 2026 fall in award value relative to Ministry of Finance spending has two competing explanations. ## T2′ — Awards move share prices, measured from the exact day the market could know (forward) Events: awards to mapped issuers whose first sighting by our collector (`tender_change.is_awarded`) is on or after 30 Jul 2026, with value/1.15 at least 1% of the issuer's market value at the previous quarter end (same definition and threshold as T2, unchanged). Event day: first trading day on or after the first sighting. Several events of one issuer within 10 trading days count once, at the first. Return: market-model abnormal return. α and β are estimated by OLS of the issuer's daily return on the TASI daily return over trading days −250 to −31 relative to the event (at least 120 valid days, else the event is dropped and counted). Primary metric: CAR over days 0 to +10. Secondary, reported but not decisive: CAR 0 to +2. Inference: bootstrap over event dates, 2,000 draws. Placebo gate: events below 0.1% of market value, same period and method. If the placebo's 95% interval excludes 0, the result is reported as "measurement not valid" whatever the main estimate. Pass: mean CAR > 0, its 95% interval excludes 0, and the placebo gate holds. When: first evaluated with the first monthly release in which at least 30 events exist; if fewer than 30 exist at the release of 3 Jul 2027, it is reported then as inconclusive with its count. Evaluated once; later reruns are labelled updates. The event threshold was not chosen with any return in view: no return of any award first seen after 30 Jul 2026 has been computed. ## T2″ — Same test, with the Tadawul disclosure date as the event day (retrospective, conditional) Events: T2's events (2019 onward, ≥1% of market value) for which the issuer published a contract-award announcement on the Saudi Exchange that names the award; event day = the announcement's trading day (or the next one, if published after the close). Method, metric, placebo gate and pass rule as T2′. Condition: run only if announcements can be matched for at least 60% of the events; otherwise reported as not run, with the match rate. Returns around announcement dates have not been computed. ## T5 — Did the 2026 fall persist in Q3? (descriptive classification) Ratio: Q3 2026 award value (award date in Jul–Sep 2026, VAT included, all awards, as T4) over Ministry of Finance Q3 2026 spending on goods and services plus non-financial assets, first published vintage. The award total is frozen on the day the Ministry publishes its Q3 2026 report (late-arriving awards after that day are not added). Reference: the Q3 ratios for 2022–2025, computed the same way. Classes: "recovered" if the ratio is at least the lowest of the four reference ratios; "persistent" if it is below 60% of their mean; "partial" otherwise. No pass/fail. A "persistent" result supports the reading that the index measures less of government purchasing in 2026, and the coverage warning on the site stays; "recovered" supports a lagged slowdown. Disclosure: the Q3 2026 award total is already visible on our site; the Ministry's Q3 2026 figure is not yet published (expected in November 2026), so the class is unknown at registration. --- # Third registration — T2‴ (registered 2026-10-30, after T2′ ran) T2′ was evaluated on 30 Oct 2026, as soon as its 30-event threshold was found already met. It passed its registered criterion, and the result is published unchanged. Diagnosed after the result (exploratory, not a re-specification): 45 of its 50 events are older awards (median 497 days between award date and first sighting) whose "first sighting" falls on 30–31 Jul 2026 because the collector began stamping that day, and the 50 events fall on only 5 distinct days (36 on 30 Jul). The pass therefore measures a market move around the start of stamping, not reaction to new awards. T2‴ is registered before any of its data is computed to fix that design flaw. ## T2‴ — Awards move share prices, fresh awards only (forward) As T2′ (threshold ≥1% of market value, market-model CAR days 0 to +10, estimation window −250 to −31, issuer de-duplication within 10 trading days, bootstrap over event dates, the same placebo gate), with two changes: - Events: first sighting on or after 1 Nov 2026, and first sighting at most 30 days after the award date (a new award, not a backlog first stamped late). - Minimum: at least 30 events on at least 20 distinct event days. When: the first monthly release meeting both minimums; if not met at the release of 3 Jan 2028, reported then as inconclusive with its counts. Evaluated once. ## T6 — Contracts running out predict re-tender awards in the same buyer × activity (forward) Registered 30 Oct 2026, from the indicator design round (lab/proposals/2026-10-30, spec `expiry-retender-pipeline`, after adversarial review). No forward data exists: the first target window opens after this commit. - Contracts: awards with a stated `contract_duration_months`, framework agreements excluded. End = award_date + 30 days + duration × 30.44 days. Annualised value AR = value_sar_excl_vat × 12 / duration. - Cell c = (merged buyer_id, activity_en). Signal X_c(t) = Σ AR of contracts known at t (award_date + 90 days, or first sighting where stamped) ending in (t, t+183d]. Renewal rate ρ_a per activity = share of expiries followed by an award in the same cell dated end−90d to end+180d worth ≥ 25% of AR; computed by script from the release of 3 Jan 2027 on expiries whose windows closed by 30 Jun 2026, and committed before the first evaluation month. ERP_c(t) = ρ_a × X_c(t). - Target Y_c(t) = Σ value_sar_excl_vat of awards in c with award_date in (t−90d, t+183d], final vintage (read 180 days after the window ends). - Months: t = month ends from 31 Jan 2027 to 31 Dec 2027 (W1; the earliest target window opens 2 Nov 2026). Confirmation W2: t = 31 Jan 2028 to 31 Dec 2028, ρ refreshed to windows closed by 30 Jun 2027. - Statistic: OLS ln(1+Y) = α_c + δ_t + β·ln(1+ERP) + Σ_{j=1..3} γ_j·ln(1+Y_c(t−365j)) + e, SE clustered by buyer. Primary: β. - Pass: β > 0 with one-sided p < 0.05 after Holm across all forward primaries registered by the evaluation date, AND within-cell ΔR² ≥ 0.02 over the same model without ERP. If β passes but the secondary β on cells where ≥ 50% of expiring value has durations outside 11–13 months does not (p ≥ 0.05), it is published as a recurrence model, not an expiry signal. - Placebo: the same regression on contracts ending (t+365d, t+548d]; β_shift must not be > 0 at 5%, and β − β_shift > 0 at 5%. - Minimum: ≥ 2,000 cell-months with X > 0 and ≥ 300 buyer clusters; else inconclusive with counts. - W1 pass is provisional; confirmed only if W2 passes. Results published as they come. Thresholds above are fixed here. - Known limits: contract start is not a field (±2-month sensitivity published); extensions are invisible; the 2026 coverage break may lower β and is reported, not excused. Deviation from the spec, stated: the spec's W1 began at t = Nov 2026, whose target window (t−90d) reaches back into awards already in the September 2026 release; W1 starts at t = Jan 2027 so no target award predates this registration. --- # Fourth registration — T7, descriptive series, attribution notes Registered: 2026-11-01, before the T7 series was computed under this definition or compared with its target. The commit that merges this section to main is the record. Disclosure of what was seen first: on 1 Nov 2026, quarterly totals of an earlier form of the contracted activity measure (tender level, no framework exclusion, missing durations set to 12 months, 2023Q1 to 2027Q4) were computed and read while describing the index. They were not compared with Ministry of Finance spending, revenue or prices. The Ministry series itself is public and was already used in T3 and T4. ## T7 — Contracted activity tracks Ministry of Finance spending (forward) - Series CAI(q): for every award known by the last day of q, value_sar_excl_vat × (days of the contract falling in q ÷ contract days), summed at tender level. Contract days start on the award date. Durations are as stated; a missing duration takes the median of its award type at the same release. Framework agreements are excluded. Known means first sighting where stamped, otherwise award date + 30 days. No single-award cap. - Target MoF(q): goods_services + capex, first published vintage in lab/targets/mof_quarterly.csv. - Statistic: Pearson correlation between g_CAI(q) = ln CAI(q) − ln CAI(q−4) and g_MoF(q) = ln MoF(q) − ln MoF(q−4), both same quarter, no lag. CAI(q−4) is taken as known on the last day of q−4 under the same rule. - Quarters: 2027Q1 to 2028Q4 (8 observations). No quarter whose awards were published before this registration is scored. - Pass: correlation > 0 with one-sided p < 0.05 after Holm across all forward primaries registered by the evaluation date. - Secondary, not decisive: the same correlation for the trailing four-quarter award value, same rule, so a pass can be read against the simpler measure. - Minimum: 8 quarters with a published target; otherwise inconclusive with counts. Evaluated once, when the 2028Q4 report is published, expected early 2029. - Holm family: T6's rule covers all forward primaries registered by its evaluation date, so T7 joins that family and makes T6's threshold stricter. This is accepted and stated here. - Known limits: 8 observations give low power; a fail is reported as "not shown", not as "no relation". Awards are commitments and the target is cash. Contract start is not a field, extensions and cancellations are invisible, and awards outside Etimad are absent. The 2026 fall in the ratio of awards to spending enters the base of 2027 growth rates and is reported, not excused. ## Descriptive series, no predictive claim DAALL:FLOW (monthly and three-month award value, year-on-year, single-award cap as in the DGI, with a per-working-day version), DAALL:CAI (T7's series on published values, 2023 average = 100, and by the eight sector groups), DAALL:EXP (annualised value of contracts ending per coming quarter, using T6's end date, by sector and largest buyers), DAALL:BREADTH (buyers increasing against decreasing, monthly). Any comparison of these with revenue, prices or spending other than T7 and T6 is labelled exploratory. ## Notes on attribution to each winner (no change to any criterion) From the release that adopts per-winner attribution, the award-level file has one row per winner of a split tender. - T1 and T2: the published results stand. A rerun at winner level is exploratory and published beside them. - T2′: already evaluated; unchanged. - T2‴: events read winner rows, so the award value is the winner's share. Thresholds and dates are unchanged. 0 events had accumulated when this note was written. - T2″: stays on the 213 events of T2. Its 60% condition is measured on those events. - T5: uses whole-tender sums, which winner rows reproduce; unchanged. - T6: stays at tender level, by summing winner rows back to tenders in signal, ρ and target, as registered. --- # Fifth registration — T8, the award queue Registered: 2026-11-02, in commit 821b8fb, before the award queue series was computed or compared with its target. That commit reached main in the same merge as the series, so the commit order on its branch is the record, not the merge. Disclosure of what was seen first: the study of the 2026 award conversion drop (October 2026, Daall strategy repository) published conversion within 180 days of closing and median days to award by closing quarter, for tenders with an Etimad identifier. On 2 Nov 2026, to fix the universe, tender counts by closing quarter, identifier, status and source were read, and days from closing to award in 2023–2024 (public tender median 78, 95th percentile 264; direct purchase 30 and 163). No count of the queue as defined here was computed, and on that day nothing was compared with award flow, revenue, prices or spending. That study itself had already set award value against Ministry of Finance spending, counted tenders stuck without an award by publication quarter, read Etimad's own monthly list totals and estimated the share of awards still missing; and the October design round (lab/proposals/2026-10-30) had specified and reviewed a tender pipeline index with award value as its target. T8's design was fixed knowing all of these. ## DAALL:AWQ — descriptive series, no predictive claim - Universe: public tenders and direct purchases with an Etimad identifier and a closing date. Rows without an identifier cannot be followed to an award; there are thousands a quarter to 2024 Q4 and few from 2025 Q1, so nothing before January 2025 is published or compared. - Not in the queue: status cancelled, or status awarded while the award record has not arrived. Cancellation is the status as read at the release, because no date exists before September 2026. - Queue(t): tenders closed in the 365 days to t with no award dated on or before t, in total, by type and by age from closing (0–90, 91–180, 181–365 days). Totals start at December 2025 and the first two bands at July 2025, when their windows hold only the comparable universe. Latest 45 days provisional. - Conversion by closing month and type: share awarded within 90 and 180 days of closing, and median days among those awarded within 180, published once the window has passed. - The latest queue by the eight sector groups. Counts only: contract value is not known before award. No buyer table. - Past points use awards dated later and today's cancellations, so they are not what could be known then. Each release is archived, so readings as known at the time accumulate from the first release. ## T8 — The award queue leads award flow (forward) - Dates: t = quarter ends from 31 Mar 2027 to 31 Dec 2028 (8 observations). - Signal: g_Q(t) = ln Queue(t) − ln Queue(t − 12 months), total queue, both read from the first release with as-of date t. Queue(t) is as known at t; the base is recomputed in that same release with hindsight, and this asymmetry is accepted and reported. - Target: g_F(t) = ln ΣFLOW(t+1..t+3) − ln ΣFLOW(t−11..t−9), DAALL:FLOW capped monthly award value, read from the first release at least 180 days after the window ends. - Statistic: Pearson correlation of g_Q and g_F. Pass: correlation > 0 with one-sided p < 0.05 after Holm across all forward primaries registered by the evaluation date. - Secondary, not decisive: the same correlation for persistence, ln ΣFLOW(t−2..t) − ln ΣFLOW(t−14..t−12), so a pass can be read against the simpler measure. And the same test for the public tender queue alone. - Minimum: 8 quarters with the target read; otherwise inconclusive with counts. Evaluated once, expected in 2029. - Holm family: T8 joins the family of T6 and T7 and makes their thresholds stricter. This is accepted and stated here. Its last target is read in about October 2029, after T7 and T6 W1 are due. They are evaluated on their own dates, counting T8 as a family member with p = 1 for ordering, and T8 is judged at its date with their p-values fixed. - Known limits: 8 observations give low power, and a fail is reported as "not shown". The sign is not obvious: a growing queue can mean more demand or slower awarding, and only the first favours this test. Counts ignore size. A change in Etimad's identifiers or in how we collect tenders is reported, not excused. ## Changes after the first computation (2026-11-02, before anything was published) An independent recomputation found that the cancellation status is almost unrecorded before September 2026: tenders it removed from the queue were 38 at December 2025 and 1,084 at October 2026, so excluding them made the queue look flat (+0.8%) when on one basis it rose (+10.1%), and the bias ran opposite to what this section first said. So, changed after the series had been seen but before any comparison with the target: - Cancelled tenders stay in the queue and in conversion, with a count of those now marked. The line above saying "Not in the queue: status cancelled" no longer holds; only "status awarded while the award record has not arrived" is left out. Conversion by month moves by at most 0.5 points in 2025 and at most 5.9 points in 2026 (direct purchases closed in July 2026). - A count of tenders left out for having no identifier is published beside the queue (788 at December 2025, 1,050 at October 2026). They are not added to the queue, so "few" above understates them: about 9% of the queue. - "Nothing before January 2025 is published" now means no queue or conversion figure. The page gives 2023–2024 days to award as reference. - The registration's other terms and all of T8's criteria are as first committed, except the three edits disclosed here: the record sentence and disclosure at the head of this section, "the first release" in the signal, and the Holm timing rule.