Tapioca Starch Outlook

Section 1

Where the forecast lands

The plateau held for four weeks, then ticked up. With a very strong El Niño now near-certain through the 2026/27 harvest, the ensemble no longer eases from September — it grinds higher into Q1.

May–December 2026 macro-regime forecast · 1–2–1 filtered with historical anchors

Observed history unchanged · Cambodia/China/Vietnam macro overlay · weak seasonality · 1–2–1 smoothing uses nearest observed anchors

Climatology ± 1σ

Observed

Ensemble starch (top 7)

Root observed

Root forecast

Ridge α=1 (best single)

GBM

Random Forest

OLS · AR(1)

Random Walk · 3MA · Climatology

Trained on monthly data through June 2026 · July nowcast anchored to TTSA 7 and 14 July · Macro overlay: border disruption, Thai root tightness, China demand and regional rerouting · Forward values 1–2–1 smoothed

        ▸ Pinch / scroll to zoom · drag to pan
        
      

Monthly forecast values All available contributors · ensemble in gold

Ensemble weighting Inverse RMSE · selected contributors from walk-forward test

Why include benchmarks. Random walk and climatology each capture something the fitted models miss. Random walk anchors the near-term level and, on this retrain, beats AR(1) and both tree models on walk-forward RMSE — a reminder that at one month ahead this series is close to a martingale. Climatology is deliberately down-weighted to 2%: at RMSE 2.97 it is worse than useless as a predictor and is retained only so the chart shows how far the current regime sits from a normal year. The macro-regime overlay carries the most weight because the binding constraints — border closure, El Niño, and Chinese demand — are structural, and no amount of history teaches a fitted model what a very strong El Niño does to a crop that has not been harvested yet.

What the next eight months look like

August–September

Take coverage now; the plateau is breaking upward, not down

July closed at 22.15 — three identical TTSA prints — and 4 August came in at 22.30 with FOB nudging 700 → 705. The ensemble carries Aug 22.25 and Sep 22.28. The previous run expected easing from September on new-crop pressure. That call is withdrawn: the crop that would have supplied that relief is the one now exposed to El Niño.

October–December

Cover Q1 2027 before the harvest print confirms it

Oct 22.36, Nov 22.46, Dec 22.56. IRI has Niño 3.4 peaking near +2.4°C in OND with 23 of 26 models at very-strong, and a positive IOD from August — the two together are the driest configuration for northeastern Thailand. That window is exactly root bulking and early harvest. Lock volume before the November root print makes the shortfall public.

January–March 2027

Expected peak — 22.62 in January

The ensemble peaks at 22.62 in January 2027, +2.1% above the July anchor, holds 22.60 in February, then turns. This is a supply peak, not a demand peak: FOB has sat at 700 USD/t for seven straight weeks, so buyers are not chasing. If Chinese chip demand softens while roots stay tight, mills — not growers — absorb the squeeze.

Through 2027

The baht has reversed — that changes the FX risk

USD/THB ran from 31.3 in February to 33.5 in July, its weakest since April 2025, before easing to 33.1. A weaker baht lifts THB export receipts and quietly supports the domestic floor — the opposite of the Q4 risk flagged in the last run. The exposure now sits the other way: if the baht retraces toward 32 while starch rolls over from the February peak, the two compound against export margin.

Section 3 · Decomposition

What you pay for, when you buy a kilo of starch

Two-thirds of the price is cassava root. The rest is what mills earn for converting it — and that spread is now narrowing as roots climb faster than starch.

Today's starch price decomposed · 22.30 THB/kg

As at 30 Jun 2026 · root midpoint @ 3.48 · μ = 6.33×

67% · ROOT MASS

11% · CONVERSION

22% · MARGIN

15.00 THB/kgCassava root × (1/η) where η ≈ 0.24
Driven by yield, planting area, weather

2.50 THB/kgEnergy, drying, packaging, labour
Driven by oil price, grid tariff, wages

4.80 THB/kgDemand-side margin
Driven by China, corn substitution

Crush margin · what a mill earns per tonne processed

Industry-standard frame, comparable to soybean crush · current values

Component	Value	Comment
Starch sell	22.30 THB/kg starch	TTSA domestic, 4 Aug 2026 · FOB 705 USD/t
Root cost (at gate)	3.60 THB/kg root	25% starch midpoint; TTTA 21 Jul range 3.20–4.10
Root mass cost in starch	15.00 THB/kg starch	3.60 ÷ η (η = 0.24)
Conversion (energy, dry, labour)	2.50 THB/kg starch	Industry standard, sticky
Crush margin	4.80 THB/kg starch	Gross spread before SG&A
Per tonne root processed	≈ 1,152 THB / tonne root	4.80 × η × 1000

A 1,000 t/day mill running 200 operating days at this spread earns ~154 M THB/year in gross margin, down from ~161 M at the June print. Floor is ~1.0 THB/kg starch (run on contracts only); break-even is ~1.5. The spread is still wide by recent standards, but the direction has turned: roots gained 4.3% in July while starch gained 0.7%. Watch the 21 July range top of 4.10 — if district medians follow it, the squeeze lands on mills, not growers.

Multiplier μ = Starch / Root · 2017–2026

When μ widens, mills earn more per kg processed

Multiplier μ (monthly)

Observed mean 2023–2026 (5.92×)

Authors' calculation · Root price hand-constructed from CEIC/OAE annual + Krungsri Research + measured Dec 2025–May 2026

What the decomposition shows — and a correction. This panel now runs on observed TTTA root prices rather than the reconstructed series used previously, and that changes the reading. On real data the compressed year was 2023 (μ = 5.43×), not 2025: roots held 3.04–3.48 while starch had not yet caught up. 2025 was in fact a comfortable year for mills (μ = 6.42×) — roots collapsed to 2.12 while starch held near 13.2, so the margin widened even as headline prices fell. The 2026 year-to-date mean is 5.85× and the current print is 6.11×, giving a crush margin of 4.80 THB/kg starch (≈ 1,152 THB/tonne root). For SMS, SQS and the rest of the Buriram–Chaiyaphum cluster the absolute spread is still healthy, but it is compressing, and an El Niño root market into Q1 2027 compresses it further. The margin story for the next four quarters is about defending μ, not riding the price.

Annual mean multiplier

2023 onward uses observed TTTA root prices · earlier years are reconstructed

Section 4

What is moving the price now

Niño-3.4

+2.1°C

▲ weekly, 15 Jul · very strong

USD/THB

33.12

▲ baht at 15-month low

CBOT Corn

$4.36

▼ off the 24 Jul high

Cassava yield

17.0

▼ El Niño risk to 2026/27

Root price

3.60

▲ +51% YoY · range top 4.10

SSE Composite

3,910

▼ −5% over July

Starch price against three reference series · annual averages

Standardised to common scale

Starch (target)

CBOT corn

Niño-3.4

Shanghai Composite

Three forces dominate, each on a different timescale. The first is corn. When American corn rallied past $7 a bushel during the Russia–Ukraine grain disruption, Asian starch buyers booked tapioca volume in advance. The displacement showed up in Thai mills not the next quarter but the next harvest — about fifteen months later. The 2023 starch high of 19 baht traced to corn prices set in late 2021.

The second force is climate, and it has changed character since the last run. What was an El Niño watch in April is now an advisory on a very strong event: Niño 3.4 averaged +0.98°C over April–June, reached +1.55°C in June, and the weekly index centred on 15 July printed +2.1°C. IRI's July plume has every one of 26 models positive through early 2027, with 23 of them at very-strong for OND. A positive IOD is forecast to establish from August and hold above 80% probability through November. El Niño with a positive IOD is the driest configuration for the northeast, and the window it covers — October to February — is root bulking and early harvest for the 2026/27 crop. The eighteen-month lag from peak Niño to peak mill price still says the deepest price impact lands in 2027 and 2028. But the front end no longer waits: buyers who can read a plume will bid for roots before the shortfall is measured.

The third force is China. The Shanghai Composite, often dismissed as retail noise, works as a proxy for Chinese industrial activity. Its rises and falls foreshadow demand for paper, adhesives and bioplastics — the dominant end-uses of Thai starch — by roughly ten months. That signal has cooled: Shanghai fell about 5% over July on a one-year low in the composite PMI and a broad reset in AI-related valuations, ending the run that had been supporting the demand-side case. It is still up on the year, but the ten-month lead now points to a softer late-2027 pull rather than a strengthening one. The model takes that seriously. It does not take the S&P 500 or Dow seriously: their lead times are weaker and the transmission less direct.

What the model does not do is treat all variables as equally informative. The cassava yield estimate, prominent in industry conversation, contributes less than corn from a year before. The dollar-baht rate, often cited as a major driver, registers modestly. The important correction is that border closure and trade rerouting are not sinusoidal. They enter as a macro-regime term that holds price level higher until confirmed supply relief appears.

News update, 7 August 2026. TTSA weekly domestic held at 22.15 for three consecutive prints (7, 14, 21 July) and moved to 22.30 on 4 August; FOB Bangkok has been pinned at 700 USD/t since mid-June and lifted only to 705. The flat FOB against a rising domestic quote says the marginal bid is coming from inside Thailand, not from export buyers. On the root side, the TTTA quoted range top jumped from 3.80 to 4.10 on 21 July while the floor stayed at 3.20 — dispersion widening upward is usually the first sign of physical scarcity, before the median moves. The TTTA ethanol reference price rose from 21.98 to 25.07 THB/litre in early July, adding a competing bid for the same roots. Krungsri's standing view remains that 2027 exports weaken on El Niño supply shortage, expanding Chinese corn supply, and Cambodian competition; the border closure keeps neighbouring feedstock out of Thai mills. The forecast treats Cambodia, China and ENSO as macro structure, not seasonality.

Section 5 · Method

How the forecast is built

Three parallel approaches. A direct multi-model ensemble for prediction, a two-equation decomposition for interpretation, and a macro-regime overlay for trade-disruption shocks.

Model horse race · walk-forward validation

▸ EXPAND

Every contributor evaluated on 67 monthly out-of-sample steps, January 2021 through July 2026, expanding-window and strictly one-step-ahead. Nothing in the training window ever sees a future value. RMSE in THB/kg.

Read this table honestly. On the extended sample the ordering is OLS 0.469, Ridge 0.501, random walk 0.584, AR(1) 0.587, gradient boost 0.664, random forest 0.920, 3-month MA 1.043, climatology 2.971. Two of these are measured here but excluded from the forward path: random forest for scoring below random walk, and AR(1) because its fitted coefficient is explosive (φ̂ = 1.024) and cannot be recursed twelve months. The linear models beat random walk by about 0.1 THB/kg — real, but a thin edge, and it is concentrated in the 2026 run-up where the corn and ENSO terms happened to point the right way. The tree models now sit below random walk: with 22 features and roughly 130 usable months they are fitting noise, and their walk-forward errors during the March–June surge were 1–2 THB/kg. The clear win remains over the published 3-month moving average baseline (Komkul, 2017^[1]), where RMSE more than halves, and over climatology, which is not a forecast at all in a regime like this one. The honest summary: at one month ahead this series is nearly a random walk, and the ensemble earns its keep at multi-month horizons and at turning points, not in the monthly error column.

The equations, written out

The two-equation decomposition makes the economic structure explicit. Root and multiplier are modelled separately, then combined. A macro-regime term is added when trade policy or border closure changes feedstock availability independently of normal seasonality.

Identity

Decomposition of starch price

\[ S_t = R_t \times \mu_t = R_t \times \left( \frac{1}{\eta} + \frac{c_t + m_t}{R_t} \right) \]

Where S_t is domestic starch price (THB/kg), R_t is cassava root farm-gate price, η ≈ 0.24 is the starch extraction yield (mass fraction), c_t is the conversion cost component (energy, drying, labour, packaging), and m_t is the residual demand-side margin. The multiplier μ_t = S_t/R_t averages 5.92× across the observed-root period Jan 2023 – Jul 2026, ranging 5.26–6.96×. Earlier years use a reconstructed root series and are not used to set the mean.

Equation 1 · Root

Physical supply equation

\[ R_t = \alpha_R + \sum_{i \in \{1,2,12\}} \phi_i R_{t-i} + \beta_1 Y_t + \beta_2 \Delta Y_t + \gamma_1 N_{t-12} + \gamma_2 N_{t-18} + \gamma_3 \bar{N}_{[t-21,t-15]} + s(t) + \epsilon^R_t \]

Root price as autoregressive in own-lags R_{t-1,2,12}, plus current and YoY-change cassava yield Y_t, ΔY_t, plus Niño-3.4 anomaly at lags 12, 18, and a 6-month average centred on lag-18. s(t) is harmonic seasonality. Coefficients estimated by Ridge regression with α=1.

Equation 2 · Multiplier

Margin equation

\[ \mu_t = \alpha_\mu + \sum_{i \in \{1,2,12\}} \psi_i \mu_{t-i} + \delta_1 C_{t-15} + \delta_2 \bar{C}_{[t-18,t-12]} + \delta_3 C_t + \theta_1 SSE_{t-9} + \omega_1 X_t + \omega_2 \Delta X_t + s(t) + \epsilon^\mu_t \]

Multiplier as autoregressive plus CBOT corn at the empirically-tuned lag-15 (substitution effect), plus contemporaneous corn, plus Shanghai Composite at lag-9 (China demand cycle), plus current and 6-month-changed USD/THB rate X_t, ΔX_t.

AR(1)

Time-series ensemble component

\[ S_t = c + \phi_1 S_{t-1} + \epsilon_t \]

First-order autoregression on starch price. One non-seasonal AR term φ_1 captures monthly persistence. The AR(1) form is the simplest pure time-series benchmark above random walk; it adds a single mean-reversion parameter relative to RW. The 2020 TTSA benchmark study^[2] used ARIMA(2,0,2); AR(1) is a parsimonious counterpart that competes with the linear regression family on the same data.

Final forecast

Macro-regime ensemble + 1–2–1 smoothing

\[ \hat{S}_t^{\text{ENS}} = \sum_{k=1}^{7} w_k \hat{S}_t^{(k)} , \quad w_k = \frac{1/\text{RMSE}_k}{\sum_j 1/\text{RMSE}_j} \]

Weights are inverse-RMSE-squared, floored at 2% and capped at 28% so no single contributor can run the forecast, then renormalised. Seven components: the macro trade-stress overlay at 29%, OLS 21%, Ridge 19%, random walk 14%, gradient boost 11%, 3-month MA 4%, climatology 2%. Two candidates are excluded and it is worth saying why. Random forest scores worse than random walk on walk-forward RMSE, which is an overfitting signal, not a contribution. AR(1) fits an explosive coefficient on this sample — φ̂ = 1.024 — so recursing it twelve months produces a monotone climb to about 25.7 THB/kg that is an artefact of a trending sample rather than a forecast. It stays in the horse race, where one-step-ahead RMSE is a fair measurement, but it is not allowed to steer the forward path. The overlay is a judgment path, not a fitted model, and its 0.36 RMSE entry is an assumed skill level rather than a measured one — it is the single most consequential assumption on this page and should be read as such. It carries the most weight because the binding constraints for the next four quarters are a border closure, an ENSO event that has not yet reached the crop, and Chinese demand, none of which the fitted models have ever seen a precedent for. Forward values are 1–2–1 smoothed twice with the July observation and the 4 August print as fixed anchors; August is blended 65/35 toward the observed 22.30 rather than taken from the models.

Lead-lag tuning

Why the lags are what they are

\[ \tau_X^* = \arg\max_\tau \left| \rho\left( \Delta X_{t-\tau}, \Delta S_t \right) \right| \quad \tau \in [0, 24] \]

Each exogenous variable's lag is selected by the cross-correlation between its three-month change and the three-month change in starch price, over horizons 3–24 months. Differencing is essential: on raw levels every one of these series correlates above 0.5 with starch simply because both trend, which would manufacture relationships that do not exist. Lags under three months are excluded because a near-contemporaneous match is not a usable lead. Result on the extended sample: τ*_Niño = 19 (r=−0.45), τ*_corn = 16 (r=+0.39), τ*_SSE = 10 (r=+0.36), τ*_USDTHB = 24 (r=+0.17). One caveat worth stating: the lag is chosen by searching 22 horizons across 7 variables, so some of the apparent precision is selection. The corn-15/16 and ENSO-18/19 results are stable across retrains and have a mechanism; the FX result does not, and should be read as noise.

Lead-lag map ▸ EXPAND

▌ Cross-correlation on three-month-differenced monthly data · positive lag = variable leads · search restricted to ≥3 months

          ▌ Niño-3.4 at 19 months captures the growing-season impact on next year's harvest — the strongest single lead-lag relationship on the board at r = −0.45. CBOT corn at 16 months reflects substitution decisions made by buyers a full crop year ahead. Shanghai Composite at 10 months leads Chinese industrial demand for paper, adhesives and bioplastics. USD/THB scores weakly and its selected lag moves between retrains; treat it as context, not signal.
        

In-sample fit · ensemble best model

One-step-ahead walk-forward predictions vs observations

Observed

Predicted

Signal strength · climate-first ranking

▌ Cross-correlation magnitude at the optimal lag for each driver

        ▌ Ranked by |r| at each variable's empirically-tuned lag — not by tree-based feature importance, which inflates AR terms. Niño-3.4 leads because cassava grows over a 12–18 month cycle and the planting decision responds to the prior wet season's rainfall, which ENSO sets a year ahead.
      

All contributors · walk-forward results

Section 6

Where this year sits in ten years of price

The seasonal arc is visible historically, but 2026 is a supply-disruption regime. Every print since March has been above the envelope, and July sits close to two standard deviations clear.

Climatological envelope · 2026 highlighted

Mean ± 1σ from 2017–2025 · 2026 traced in sky blue

Climatological mean ± 1σ

Climatological mean

2026

Where July 2026 lands. Climatology mean for July is 16.23 with σ ≈ 3.10 — note that σ has widened as 2026 enters the reference sample, which mechanically shrinks the apparent anomaly. The July mean of 22.15 sits at +1.91σ, the highest July on record for this series and above the 2023 July peak of 18.50. May was revised down from 19.50 to 18.98 in this build: the earlier figure had taken a single mid-month print rather than the mean of the four weekly quotes. All 2026 monthly values are now straight means of the TTSA weekly prints. The 2025 trough is well behind us, and the forecast treats 2026–27 as a supply-constrained regime rather than a plateau waiting to unwind.

Departure from climatology · year by year

Each line: that year's biweekly price minus long-term mean

Reading the year-by-year departures. The last ten years split into three regimes. Below climatology: 2019, 2020, 2021, 2025 — buyer-friendly years with weak corn pull-through and ample root supply. At climatology: 2017, 2018. Above climatology: 2022, 2023, 2024 — the post-Ukraine corn squeeze that lifted Thai starch by 4–5 THB/kg through 2023 and held it elevated for two years. 2026 is now the most extreme member of the third group. July prints +5.9 THB/kg above the long-run July mean, roughly double the largest departure of the 2022–24 episode. Neither historical analogue fits: 2022 peaked in July and plateaued, 2024 peaked in April and fell through Q3, and both were demand-led. This one is feedstock-led with a very strong El Niño still ahead of the harvest, which is why the forecast declines to impose either shape and leans on the macro overlay instead.

Literature on Thai cassava price and export forecasting

Komkul, P. (2017). Forecasting Cassava Starch Price in Thailand by Using Time Series Models. The Journal of King Mongkut's University of Technology North Bangkok. Compared Box-Jenkins, Holt's exponential smoothing, damped trend, and 3/6/12-month moving averages on TTSA monthly data 2009–2014; the 3-month moving average had the lowest MAPE for 2014.
Anonymous (2020). The impact of declining export price of tapioca starch on Thai economic output and employment. ResearchGate working paper. ARIMA(2,0,2) selected by minimum AIC on TTSA monthly 2011–2020; 8-month-ahead forecast.
Pannakkong, W., Huynh, V.-N., & Sriboonchitta, S. (2016). ARIMA Versus Artificial Neural Network for Thailand's Cassava Starch Export Forecasting. In Causal Inference in Econometrics, Springer. ANN models outperformed ARIMA on MSE/MAE/MAPE for native starch, modified starch, and sago export volumes 2001–2013.
Pannakkong, W., Sriboonchitta, S., & Huynh, V.-N. (2019). A Novel Hybrid Autoregressive Integrated Moving Average and Artificial Neural Network Model for Cassava Export Forecasting. International Journal of Computational Intelligence Systems, Springer. Hybrid ARIMA-ANN model with seasonal index gave the lowest error on native and modified starch export forecasts.
FAO (2002). Forecast of area, yield and production of Thai cassava roots. Paper 5, FAO/OAE/Kasetsart econometric study. Planted-area function on lagged own price, competing crop prices, and lagged area; yield by time-series moving average; OLS-estimated Cobb-Douglas at national and regional levels.
Sriroth, K. (1999). Cassava industry in Thailand: The status of technology and utilization. International Symposium on Cassava, Starch, and Starch Derivatives, Nanning, China. Background reference on starch extraction yield (η ≈ 23–25%) used in the decomposition above.

Distinction from the published literature. All five forecasting papers cited above are univariate (price or export volume forecast from its own past). The model presented here adds exogenous regressors at empirically-tuned lags — a methodologically more ambitious approach not previously documented for Thai tapioca starch.

KUCCI/Cassava Price Terminal

FEED · waiting for cassava_feed.js

Roots 25% TTSA Kasetprice farm-gate NETTA R25 NETTA R30 R25 fcst Starch TTSA Starch fcst

OPERATIONAL SIGNALS / NETTA · DAILY

NETTA daily · last district snapshot 18 May

Chips / Ayutthayaมันเส้น

—THB/kg

—

Upstream signal for starch — when chip price rises, roots get diverted from starch processing, tightening starch supply with 2–4 wk lag.

Dry Residue / Mueangกากมัน

—THB/kg

—

Post-extraction byproduct sold as feed. Rising price corroborates firm feed-market demand → competitive bid for chips.

Industry CapacityUTILISATION

<50%

unchanged · ~5 wk flat

Factories running below half capacity. When this flips above 50%, supply has loosened. Currently supply-constrained. The TTTA ethanol reference price rose 21.98 → 25.07 THB/litre in early July, adding a competing bid for the same roots.

Method · TTSA 4 Aug + TTTA 21/23 Jul + NETTA/Kasetprice anchors · level-shift + constrained-supply forecast

5-week trend-carry · TTSA weekly through 4 Aug · TTTA root range 3.20–4.10 · factory-pressure index

⌄

Observed price arrays are kept as source records and are not back-adjusted. The 23 Jul TTTA starch range (20.80–21.20) is retained as a market cross-check; the TTSA 4 Aug domestic quote of 22.30 is the weekly model anchor. Note there is no TTSA print for 28 July — the published series steps from 21 July to 4 August, so the near-term carry spans a two-week gap rather than one. Forecast = confirmed level shift + persistent supply-pressure carry + weak seasonality + median-confirmed root momentum. Seasonal reversion is capped and cannot force a downturn until root medians weaken. R25 forecast is plotted with its uncertainty spread. R30 forward calculation is retained internally and folded into the starch forecast cone instead of being plotted as a separate line.

Stage 1 — weekly baseline

$$\hat{P}^{base}_{j,h} = \hat{P}_{j,h-1} + \rho(T_{j,h}-\hat{P}_{j,h-1}) + d_h m_j$$ $$T_{j,h}=A_j\frac{1+H_j(w_t+h)}{1+H_j(w_t)}, \qquad A_j=0.70P_{j,t}+0.30\overline{P_j}^{(4w)}$$ $$\rho=0.22,\qquad d_h=[0.65,0.35,0.12,0.05,0.02]$$

Muted seasonal harmonic

$$H_R(w)=-0.020\sin\left(\frac{2\pi w}{52}\right)+0.016\cos\left(\frac{2\pi w}{52}\right)$$ $$H_S(w)=-0.012\sin\left(\frac{2\pi w}{52}\right)+0.006\cos\left(\frac{2\pi w}{52}\right)$$

Stage 2 — NETTA bias correction

$$Bias^{NETTA}=\overline{R^{NETTA}_{25}}_{(t_a,t_a+7d]}-\hat{R}^{base}_{25,1}$$ $$\hat{R}^{bias}_{25,h}=\hat{R}^{base}_{25,h}+\gamma_h Bias^{NETTA},\qquad \gamma_h=[0.85,0.55,0.30,0.15,0.05]$$

Stage 3 — factory-pressure index

$$PI_t= 0.30Z(R^{NETTA}_{30}-R^{NETTA}_{25}) +0.30Z(C^{NETTA}/2.5-R^{NETTA}_{25}) +0.15Z(G^{NETTA}) +0.15Z(R^{NETTA}_{25}-R^{weekly}_{25}) +0.20B_t +0.10U_t$$ $$B_t=+1\ \text{for constrained cross-border/import flow},\qquad U_t=0.75\ \text{when utilization}<50\%$$

Final forecasts

$$\hat{R}_{25,h}=\hat{R}^{base}_{25,h}+\gamma_h Bias^{NETTA}+\beta_h PI_t+\kappa_h(K_{25,t}-R^{NETTA}_{25,t})$$ $$\beta_h=[0.055,0.040,0.025,0.015,0.005],\qquad \kappa_h=[0.35,0.22,0.12,0.06,0.02]$$ $$\hat{R}_{30,h}=\hat{R}_{25,h}+\bar{Q}+\eta_h(Q_t-\bar{Q}),\qquad Q_t=R^{NETTA}_{30,t}-R^{NETTA}_{25,t}$$ $$\eta_h=[0.65,0.45,0.28,0.15,0.06]$$ $$\hat{S}_{h}=\hat{S}^{base}_{h}+\lambda_h PI_t+\omega_h J_t$$ $$\lambda_h=[0.105,0.075,0.050,0.030,0.015],\qquad \omega_h=[0.28,0.15,0.07,0.03,0.00]$$ $$J_t=\max\left(0,\Delta S_t-\overline{\Delta S}_{prior}\right)$$

Conversion checks

$$R^{chips}_t=\frac{C^{NETTA}_t}{2.5}$$ $$R^{starch,gross}_t=0.20S^{TTSA}_t$$